[llvm-dev] [RFC] Implementing the sanitizer runtimes for heterogeneous devices
Sharma, Reshabh Kumar via llvm-dev
llvm-dev at lists.llvm.org
Mon May 10 08:30:52 PDT 2021
Following approach #2 from the previous RFC, we are preparing some updates to the address sanitizer
runtime and we would like to get some feedback on how to approach these changes.
In order to isolate the impact of our changes on others, we are thinking to guard our changes inside a
distinguishable macro, for example, SANITIZER_AMDGPU, and to localize our changes as much as possible
within new files whose name includes AMDGPU. We would your feedback on this, and suggestions
about how to trigger the definition of the macro. The host runtime will always be targeted to the host
which implies that we will not be able to check the target architecture and set the macro, like other
architectures. We propose two ways to achieve this:
1. Based on LLVM_TARGETS_TO_BUILD value
We can look for AMDGPU in the LLVM_TARGETS_TO_BUILD string and trigger the AMDGPU
specific macro. We will make sure it does not trigger with the default option which is the list of
all the supported architectures.
1. Adding a new cmake flag
We could require a specific cmake variable be set which when used will trigger the AMDGPU
We are definitely open to other alternatives.
From: Sharma, Reshabh Kumar
Sent: Thursday, October 29, 2020 9:34 PM
To: 'llvm-dev at lists.llvm.org' <llvm-dev at lists.llvm.org>
Cc: Sumner, Brian <Brian.Sumner at amd.com>
Subject: [RFC] Implementing the sanitizer runtimes for heterogeneous devices
We are thinking of extending the LLVM sanitizers, starting with asan, to heterogeneous situations
such as those found in OpenCL and HIP. We have already started thinking about the way
instrumentation will look like for different address spaces and posted a small RFC
(http://lists.llvm.org/pipermail/llvm-dev/2020-July/143310.html) about removing the generated
inttoptr and ptrtoint pair from instrumented IR. We have started to look into the runtime
implementation, and we will be happy to get feedback and suggestion from the community so that we
can incorporate them in the design itself. We are looking at implementing a sanitizer runtime that
would support heterogeneous devices such as AMD GPUs and we present some directions that we can take
to make that happen:
1. Add support for compilation of the compiler-rt sanitizer runtime by a device compiler
We propose to update the sanitizer parts of compiler-rt so they can additionally be compiled
with the HIP compiler and generate device code. This would be our first choice. It provides
obvious advantages over keeping separate sanitizer runtimes for heterogeneous devices and
easier maintenance while ensuring feature parity. This can be broadly summarized in the
a. Supporting device code generation
We propose to add suitably protected changes which allow the code to be compiled for the
HIP language and generate device code.
b. Supporting device and host communication
Sanitizer features like violation reporting requires communication from device to host.
We propose to add support for the device code to communicate with the host for
reporting. We see this going into a target specific sub-folder.
c. Adding new interceptors
We foresee a need for new interceptors, for example, device specific memory allocators.
We propose to add these interceptors.
d. Updating the build system
We propose to update the build system to build device code when requested, as well as
2. Implement the device side sanitizer runtime separately
In this scenario, the device side of the sanitizer runtime would be implemented in a
separate folder, possibly in a separate repository. For example, this could be added to our
existing device side runtime. The changes mentioned in 1b and 1c to the existing host
runtime would still be required.
Combinations of different approaches listed above are also possible. We will be happy to hear about
other approaches and ideas to move forward with them as well. Adding direct support for
heterogeneous device sanitizer runtime in compiler-rt will also be helpful for other heterogeneous
devices in future.
Reshabh and Brian
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the llvm-dev