[CUDA] work around more __noinline__ conflicts with libc++

[Artem-B]

#73838

[llvmbot]

@llvm/pr-subscribers-backend-x86

@llvm/pr-subscribers-clang

Author: Artem Belevich (Artem-B)

Changes

#73838

---

Full diff: https://github.com/llvm/llvm-project/pull/74123.diff

3 Files Affected:

• (modified) clang/lib/Headers/CMakeLists.txt (+2)
• (added) clang/lib/Headers/cuda_wrappers/__config (+10)
• (added) clang/lib/Headers/cuda_wrappers/string (+10)

diff --git a/clang/lib/Headers/CMakeLists.txt b/clang/lib/Headers/CMakeLists.txt
index fdd54c05eedf825..f562c354327f2fb 100644
--- a/clang/lib/Headers/CMakeLists.txt
+++ b/clang/lib/Headers/CMakeLists.txt
@@ -283,10 +283,12 @@ set(files
   )
 
 set(cuda_wrapper_files
+  cuda_wrappers/__config
   cuda_wrappers/algorithm
   cuda_wrappers/cmath
   cuda_wrappers/complex
   cuda_wrappers/new
+  cuda_wrappers/string
 )
 
 set(cuda_wrapper_bits_files
diff --git a/clang/lib/Headers/cuda_wrappers/__config b/clang/lib/Headers/cuda_wrappers/__config
new file mode 100644
index 000000000000000..04038b0fa3437ee
--- /dev/null
+++ b/clang/lib/Headers/cuda_wrappers/__config
@@ -0,0 +1,10 @@
+// CUDA headers define __noinline__ which interferes with libc++'s use of
+// `__attribute((__noinline__))`. In order to avoid compilation error,
+// temporarily unset __noinline__ when we include affected libstdc++ header.
+
+#pragma push_macro("__noinline__")
+#undef __noinline__
+#define __noinline__ __noinline__
+#include_next "__config"
+
+#pragma pop_macro("__noinline__")
diff --git a/clang/lib/Headers/cuda_wrappers/string b/clang/lib/Headers/cuda_wrappers/string
new file mode 100644
index 000000000000000..d612710393623b5
--- /dev/null
+++ b/clang/lib/Headers/cuda_wrappers/string
@@ -0,0 +1,10 @@
+// CUDA headers define __noinline__ which interferes with libc++'s use of
+// `__attribute((__noinline__))`. In order to avoid compilation error,
+// temporarily unset __noinline__ when we include affected libstdc++ header.
+
+#pragma push_macro("__noinline__")
+#undef __noinline__
+#define __noinline__ __noinline__
+#include_next "string"
+
+#pragma pop_macro("__noinline__")

[gribozavr]

I don't mind this, but should libc++ start using _LIBCPP_NOINLINE more, we would be playing a catch-up game here - we would need to add more and more similar wrapper headers.

[Artem-B]

Yes, I've mentioned that in #73838. However, we need something to fix the issue right now while we're figuring out a better solution.

In any case __noinline__ is unlikely to be widely used, so the wrappers may be manageable, at least for the short-to-medium term.

[philnik777]

I think we can find a solution to work around this in libc++ within a reasonable timeframe, but I won't block this, since I don't really care about this part of the code base.

[Artem-B]

I think we can find a solution to work around this in libc++ within a reasonable timeframe

OK. I'll hold off on landing the patch. I believe we're not blocked on it at the moment.

[ldionne]

Some review comments if we end up merging this.

FWIW I am not thrilled about using __config here. That header is an implementation detail of libc++ and defining it and relying on it is somewhat brittle.

[Artem-B]

FWIW I am not thrilled about using __config here. That header is an implementation detail of libc++ and defining it and relying on it is somewhat brittle.

I'm all for having it fixed in libc++ or in CUDA SDK. Barring that, working around the specific implementation details is all I can do here. I'm open to other ideas on how this conflict can be worked around or solved in a better way.

[gribozavr]

#73838 has landed, we can close this PR now.