Fixed two issues:
1. assertion with -flto. the linker wrapper action is missing for
wrapping the device binary. Added it for -flto.
2. when there are two HIP files, the kernels in the second file were not
found. This is because the -r option of linker wrapper assumes offload
entries section of HIP to be hip_offloading_entries but it is actually
llvm_offload_entries, causing the offload entries sections not made
unique for different object files. Fixed and tested working for both
-fgpu-rdc and -fno-gpu-rdc case with and without -r
The HLSL driver currently defaults the triple to an unversioned os and
subarch when targeting SPIR-V. This means the SPIR-V backend decides the
default value. That is not a great option because a change the backend
could cause a change in Clang.
Now that we want to choose the default we need to consider the best
option. DXC currently defaults to Vulkan1.0. We are planning on not
supporting Vulkan1.0 in the Clang HLSL compiler because it is newer
versions of Vulkan are commonly supported on nearly all hardware, so
users do not use it.
Since we have to change from DXC anyway, we are using VK1.3. It has been
out long enough to be commonly available, and the initial implementation
of SPIR-V features for HLSL are assuming Vulkan 1.3.
---------
Co-authored-by: Nathan Gauër <github@keenuts.net>
Visual Studio has an argument to ignore all PCH related switches. clang-cl has also support option /Y-. Having the same option in clang would be helpful. This commit is to add support for ignoring PCH options (-ignore-pch).
The commit includes:
1. Implement -ignore-pch as a Driver option.
2. Add a Driver test and a PCH test.
3. Add a section of -ignore-pch to user manual.
4. Add a release note for the new option '-ignore-pch'.
Code reviewed by: Matheus Izvekov <mizvekov@gmail.com>
This breaks a lot of new driver HIP compilation. We should probably
revert this for now until we can make a fixed version.
```c++
static __global__ void print() { printf("%s\n", "foo"); }
void b();
int main() {
hipLaunchKernelGGL(print, dim3(1), dim3(1), 0, 0);
auto y = hipDeviceSynchronize();
b();
}
```
```c++
static __global__ void print() { printf("%s\n", "bar"); }
void b() {
hipLaunchKernelGGL(print, dim3(1), dim3(1), 0, 0);
auto y = hipDeviceSynchronize();
}
```
```console
$ clang++ a.hip b.hip --offload-arch=gfx1030 --offload-new-driver
$ ./a.out
foo
foo
```
```console
$ clang++ a.hip b.hip --offload-arch=gfx1030 --offload-new-driver -flto
<crash>
```
This reverts commit d54c28b9c1396fa92d9347ac1135da7907121cb8.
Summary:
This reverts commit 740e6aeccdb2f8eeee549ba80b15ff3d5dd3392e.
After discussions it was determined that the behavior for IR inputs
needs to be maintained at least for now. In the future we should put a
deprecation notice on this behavior. This patch keeps the old behavior
for OpenCL and IR inputs, while others will be standalone. This is good
enough for standard compile flows.
This simply moves code for diagnosing misuse of arguments `/Fo`, `/Fa`,
and `/o` from `Driver::BuildActions` into `Driver::handleArguments`,
following the intention of 740f69b.
This change better aligns with the roles of `BuildActions` and
`handleArguments`.
This PR add functionality to change `flang` command line using
environment variable `FCC_OVERRIDE_OPTIONS`. It is quite similar to what
`CCC_OVERRIDE_OPTIONS` does for clang.
The `FCC_OVERRIDE_OPTIONS` seemed like the most obvious name to me but I
am open to other ideas. The `applyOverrideOptions` now takes an extra
argument that is the name of the environment variable. Previously
`CCC_OVERRIDE_OPTIONS` was hardcoded.
When using -no-integrated-cpp, before, the driver won't collapse actions
when the input was not llvm-ir
or it would collapse them too aggressively with -save-temps
The original code was checking the action type (which is IR too for
preprocessed->bc actions) instead of the action inputs.
Add a new Cygwin toolchain that just goes through the motions to
initialize the Generic_GCC base properly. This allows removing some old,
almost certainly wrong hard-coded paths from Lex/InitHeaderSearch.cpp.
MSYS2 (GCC triple (arch)-pc-msys) is a fork of Cygwin (GCC triple
(arch)-pc-cygwin), and this driver can be used for either.
Add a simple test for this driver.
Currently HIP still uses offload bundler for non-rdc mode for the new
offload driver.
This patch switches to use offload wrapper for non-device-only non-rdc
mode when new offload driver is enabled.
This makes the rdc and non-rdc compilation more consistent and speeds up
compilation since the offload wrapper supports parallel compilation for
different GPU arch's.
It is implemented by adding a linker wrapper action for each assemble
action of input file. Linker wrapper action differentiates this special
type of work vs normal linker wrapper work by the fle type. This type of
work results in object instead of image. The linker wrapper adds "-r"
for it and only includes the object file as input, not the host
libraries.
For device-only non-RDC mode, the new driver keeps the original
behavior.
Due to `amdgcnspirv` piggybacking on the HIPAMDToolchain, it loses its
immediately apparent SPIR-Vness, which makes the Driver want to go all
the way to Assembly emmission. This was problematic as we were trying to
`llvm-link` SPIR-V, before trying to translate it. This patch ensures
that we do the right thing and stop at bitcode emission if we are mixing
`amdgcnspirv` and concrete targets.
Summary:
We need a toolchain to get the GPU architectures when compiling with
OpenMP. This kind of breaks the toolchain model because these are cached
all over the place. Instead of making a new one, just create both of
them unconditionally. It's not like this is saving any work since we
still needed to create both toolchains in the earlier case.
Fixes: https://github.com/llvm/llvm-project/issues/131325
This updates the DXV and Metal Converter actions to properly use
temporary files created by the driver. I've abstracted away a check to
determine if an action is the last in the sequence because we may have
between 1 and 3 actions depending on the arguments and environment.
This patch fixes the buildbots failure of lit tests on MacOS. Since
clang driver options depend on toolchain, we cannot hardcode CUID hash.
On MacOS there is an extra -mlinker-version= option.
Add an option similar to the -qtarget option in XL to allow the user to
say they want to be able to run the generated program on an older
version of the LE environment. This option will do two things:
- set the `__TARGET_LIBS` macro so the system headers exclude newer
interfaces when targeting older environments
- set the arch level to match the minimum arch level for that older
version of LE. It doesn't happen right now since all of the supported LE
versions have a the same minimum ach level. So the option doesn't change
this yet.
The user can specify three different kinds of arguments:
1. -mzos-target=zosv*V*r*R* - where V & R are the version and release
2. -mzos-target=0x4vrrmmmm - v, r, m, p are the hex values for the
version, release, and modlevel
3. -mzos-target=current - uses the latest version of LE the system
headers have support for
Summary:
This patch cleans up how we query the offloading toolchain. We create a
single that is more similar to the existing `getToolChain` driver
function and make all the offloading handlers use it.
Summary:
Currently, `-Xarch_` is handled specially between different toolchains,
(i.e. Mach-O).
This patch unifies the handling so that it can be used generically.
The main benefit here is that we now have a more generic version of
`-Xopenmp-target=`, which should probably just be deprecated.
Additionally, it allows us to specially pass arguments to different
architectures for offloading.
This patch is done in preparation for making selecting offloading
toolchains more generic, this will be helpful while people are moving
toward compile jobs that include multiple toolchains (SPIR-V, AMDGCN,
NVPTX).
Summary:
OpenMP was weirdly split between using the bound architecture from
`--offload-arch=` and the old `-march=` option which only worked for
single jobs. This patch removes that special handling. The main benefit
here is that we can now use `getToolchainArgs` without it throwing an
error.
I'm assuming SYCL doesn't care about this because they don't use an
architecture.
In the discussion around #116792, @rjmccall mentioned that ARCMigrate
has been obsoleted and that we could go ahead and remove it from Clang,
so this patch does just that.
The only architecture currently being supported (still a WIP) is
x86_64. Other UEFI triples targeting other architectures will now
report an `unknown target triple` error.
Summary:
The `getToolChain` pass uses the triple to determine which toolchain to
create. Currently the `amdgcn-amd-amdhsa` triple maps to the
`ROCmToolChain` which uses things expected to be provided by `ROCm`.
This is neded for OpenCL, but directly targeting C++ does not want this
since it's primarily being used for creating GPU runtime code. As far as
I know I'm the only user of this, so this shouldn't change anything.
Unfortunately, there's no good logic for detercting this, so I simply
checked ahead of time if the input is either `foo.cl` or `-x cl foo.c`
to choose between the two. This allows us to use the AMDGPU target
normally, as otherwise it will error without passing `-nogpulib`.
This patch is the third step to extend the current multilib system to
support the selection of library variants which do not correspond to
existing command-line options.
Proposal can be found in
https://discourse.llvm.org/t/rfc-multilib-custom-flags/81058
The multilib mechanism supports libraries that target code generation or
language options such as --target, -mcpu, -mfpu, -mbranch-protection.
However, some library variants are particular to features that do not
correspond to any command-line options. Examples include variants for
multithreading and semihosting.
This work introduces a way to instruct the multilib system to consider
these features in library selection. This particular patch is comprised
of the core processing of these flags.
- Custom flags in the command-line are read and forwarded to the
multilib system. If multiple flag values are present for the same flag
declaration, the last one wins. Default flag values are inserted for
flag declarations for which no value was given.
- Feed `MacroDefines` back into the driver. Each item `<string>` in the
`MacroDefines` list is formatted as `-D<string>`.
Library variants should list their requirement on one or more custom
flags like they do for any other flag. The new command-line option is
passed as-is to the multilib system, therefore it should be listed in
the format `-fmultilib-flag=<str>`.
Moreover, a variant that does not specify a requirement on any
particular flag can be matched against any value of that flag.
If the user specifies `-fmultilib-flag=<name>` with a name that is
invalid, but close enough to any valid flag value name in terms of edit
distance, a suggesting error is shown:
```
error: unsupported option '-fmultilib-flag=invalidname'; did you mean '-fmultilib-flag=validname'?
```
The candidate with the smallest edit distance is chosen for the
suggestion, up to a certain maximum value (implementation detail), after
which a non-suggesting error is shown instead:
```
error: unsupported option '-fmultilib-flag=invalidname'
```
CUID is needed by CUDA/HIP for supporting accessing static device
variables in host function.
Currently CUID is only supported by the old driver for CUDA/HIP. The new
driver does not support it, which causes CUDA/HIP programs using static
device variables in host functions to fail with the new driver for
CUDA/HIP.
This patch refactors the CUID support in the old driver so that CUID is
supported by both the old and the new drivers for CUDA/HIP.
Given libstdc++ has landed std module, the build systems may need clang
to find the configuration file to understand how to build the std
module. This patch did this. Tested with locally installed GCC-trunk.
This commit fixes -print-library-module-manifest-path on macos.
Currently, this only works on linux systems. This is because on macos
systems the library and header files are installed in a different
location. The module manifest is next to the libraries and the search
function was not looking in both places. There is also a test included.
In DXC, setting the vulkan version automatically sets the target spir-v
version to the maximum spir-v version that the vulkan version must
support. So for Vulkan 1.2, we set the spir-v version to spirv 1.5
because every implementation of Vulkan 1.2 must support spirv 1.5, but
not spir-v 1.6.
Embedded development often needs to use a different C standard library,
replacing the existing one normally passed as -internal-externc-isystem.
This works fine for an apple-macos target, but apple-none-macho doesn't
work because the MachO driver doesn't implement
AddClangSystemIncludeArgs to add the resource directory as
-internal-isystem like most other drivers do. Move most of the search
path logic from Darwin and DarwinClang down into an AppleMachO toolchain
between the MachO and Darwin toolchains.
Also define __MACH__ for apple-none-macho, as Swift expects all MachO
targets to have that defined.
On z/OS, the location of the system libraries and side decks (aka
equivalent to libc, etc) are not in a predefined location. The system
does have a default location but sysadmins can change this and
frequently do. See the -mzos-hlq* options we have for z/OS.
To avoid every user needing to specify these -mzos-hlq* options, we
added support for a system install default config file that is always
read independent of the usual config file. The compiler will read this
customization config file before reading the usual config files.
The customization file is called clang.cfg and is located in:
- the etc dir within the compiler installation dir.
- or specified by the CLANG_CONFIG_PATH env var. This env var can either
be a directory or the fill path name of the file.
Embedded development often needs to use a different C standard library,
replacing the existing one normally passed as -internal-externc-isystem.
This works fine for an apple-macos target, but apple-none-macho doesn't
work because the MachO driver doesn't implement
AddClangSystemIncludeArgs to add the resource directory as
-internal-isystem like most other drivers do. Move most of the search
path logic from Darwin and DarwinClang down into an AppleMachO toolchain
between the MachO and Darwin toolchains.
Also define \_\_MACH__ for apple-none-macho, as Swift expects all MachO
targets to have that defined.
Introduces the SYCL based toolchain and initial toolchain construction
when using the '-fsycl' option. This option will enable SYCL based
offloading, creating a SPIR-V based IR file packaged into the compiled
host object.
This includes early support for creating the host/device object using
the new offloading model. The device object is created using the
spir64-unknown-unknown target triple.
New/Updated Options:
-fsycl Enables SYCL offloading for host and device
-fsycl-device-only
Enables device only compilation for SYCL
-fsycl-host-only
Enables host only compilation for SYCL
RFC Reference:
https://discourse.llvm.org/t/rfc-sycl-driver-enhancements/74092
This is a reland of: https://github.com/llvm/llvm-project/pull/107493
This is the first of a series of patches to add support for OpenMP
offloading to SPIR-V through liboffload with the first intended target
being Intel GPUs. This patch implements the basic driver and
`clang-linker-wrapper` work for JIT mode. There are still many missing
pieces, so this is not yet usable.
We introduce `spirv64-intel-unknown` as the only currently supported
triple. The user-facing argument to enable offloading will be `-fopenmp
-fopenmp-targets=spirv64-intel`
Add a new `SPIRVOpenMPToolChain` toolchain based on the existing general
SPIR-V toolchain which will call all the required SPIR-V tools (and
eventually the SPIR-V backend) as well as add the corresponding device
RTL as an argument to the linker.
We can't get through the front end consistently yet, so it's difficult
to add any LIT tests that execute any tools, but front end changes are
planned very shortly, and then we can add those tests.
---------
Signed-off-by: Sarnie, Nick <nick.sarnie@intel.com>
Currently, the config file system loads the full target triple, e.g.
`arm64-apple-darwin23.6.0.cfg`.
This is however not very useful as this is a moving target. In the case
of macOS, that target moves every ~2 months.
We can improve this by adding fallbacks that simplify the version
component of the triple. This pull request adds support for loading
`arm64-apple-darwin23.cfg` and `arm64-apple-darwin.cfg`. See the
included test for a demonstration on how it works.
The only difference is the usage of `JobAction* JA` versus `Action* A`
in one argument, but `JA = cast<JobAction>(A)`, and the called function
is inherited from `Action`.
Currently, if a -l (or -Wl,) flag is added into a config file
(e.g. clang.cfg), it is situated before any object file in the
effective command line. If the library requested by given -l flag is
static, its symbols will not be made visible to any of the object
files provided by the user. Also, the presence of any of the linker
flags in a config file confuses the driver whenever the user invokes
clang without any parameters (see issue #67209).
This patch attempts to solve both of the problems, by allowing a split
of the arguments list into two parts. The head part of the list will
be used as before, but the tail part will be appended after the
command line flags provided by the user and only when it is known
that the linking should occur. The $-prefixed arguments will be added
to the tail part.
