I had to revert https://github.com/llvm/llvm-project/pull/138205 in
https://github.com/llvm/llvm-project/pull/141994 because it broke the
Chrome build.
The problem came down to the following:
```c++
unsigned __int64 _umul128(unsigned __int64, unsigned __int64,
unsigned __int64 *);
namespace {}
#pragma intrinsic(_umul128)
void foo() {
unsigned __int64 carry;
unsigned __int64 low = _umul128(0, 0, &carry);
}
```
When processing the `#pragma intrinsic` line, we do a name lookup to see
if the builtin was previously declared. In this case the lookup fails
because the current namespace of the parser and sema is the above
namespace scope. The processing of the pragma happens as part of the
namespace close parsing. This is usually fine because most pragmas don't
care about scopes. However, that's not true for this and other MS
pragmas.
To fix this, we change the `#pragma intrinsic` processing to be the same
as other MS pragmas such as "optimize". Those are processed like a
declaration, and because of that we have the correct current scope, so
the lookup succeeds.
I added a test case that locks down the Chrome fix, as well as manually
tested the Chrome build and confirmed it passed.
When trying to remove the usage of `__has_builtin` on MSVC CUDA ARM for
some builtins, the recommended direction was to universally declare the
MSVC builtins on all platforms and require the header providing
declarations to be included. This was done
[here](https://github.com/llvm/llvm-project/pull/128222).
However, some MSVC headers already use the MSVC builtins without
including the header, so we introduce a warning for anyone compiling
with MSVC for this target, so the above change had to be reverted.
The MSVC headers use `#pragma intrinsic` before the intrinsic uses and
that seems to be enough for MSVC, so declare builtins when used in
`#pragma intrinsic` in Clang to prevent the warning.
---------
Signed-off-by: Sarnie, Nick <nick.sarnie@intel.com>
