LGTM. Thanks Alexey.

Is it possible to have the same problem with other directives, like 'omp single' or 'omp master'? I haven't seen any test fail for those yet though.

This looks like the right change to me.

My reasoning is that the generic TreeTransform does semantic analysis on CallExpr and this seems to be related semantic analysis that would need to be redone when the call changed as well. Say someone codes up a new subclass of TreeTransform that maybe changes one of the arguments. If ActOnOpenMPCall is not called does that create a broken AST?

LGTM. I did run the math_cpp11 tests. I don't have a cuda environment so it was probably of limited value. I didn't see any failures that seemed related to this during the compiles.

Looks okay to me.

I just moved your lookup code and tried to get the same info from the declarator. The function looks like this:

For the most part, these headers are going to be included in almost every compilation unit anyway since they will come in through other C/C++ library headers. So this would presumably affect only a small numbers of compilation units that only include stddef.h and none of the others.

Added comments as requested.

Updated to use two options: -pch-through-header-create and -pch-through-header-use.

Thanks for the review. Updated based on comments from Hans.

I've made good progress on this. My lit tests are working. I will try it on the tests in our test suites and maybe I will have a patch for review in a few days.

In D50640#1203314, @thakis wrote:

Thanks! Do you need someone to land this?

In D50640#1201216, @thakis wrote:

How does the gcc driver codepath handle this?

Added a -verify test to ensure no warnings on successful PCH use.

In D46652#1164010, @thakis wrote:

Also, were you planning on also adding support for the (filename-less version of) hdrstop pragma? After this change, that should probably be fairly straightforward.

Thanks for the review. Updated based on comments.

Ping. This can significantly improve compile time for projects that use the common /Yc<header.h> /Yu<header.h> PCH mechanism.

Ping. Still looking for a reviewer, mostly Lex and clang-cl driver changes.

Sorry for the delay in getting back to this. I've found that we are using many OpenMP directives not just simd. We'd like to continue doing that.

In D46652#1101952, @kimgr wrote:

Added warning when macro is defined when skipping. Also fixed skipping so directives are completely ignored expect for #include and #define.

Thanks for taking a look at the patch. I added a warning as suggested. I also fixed skipping so it ignores all directives except #include and now #define.

1. It probably makes sense to allow omp simd with OpenCL. Some people here have been successfully using it anyway.
2. I can give that a try.