LGTM; thanks! The IBM OpenMP runtime sets the default thread stack size to 4MiB on AIX as well.

Dropped the static_assert on the size of struct __long.

Addressed the comment to consider potential padding in the assert added for struct __long.

The build failure is unrelated, can you rebase before reuploading this patch?

Addressed comments.

• add comments explaining why attribute __packed__ is used;
• add static_assert for struct __long.

Include context.

Changed to use anonymous struct and __attribute__((__packed__)) as suggested.

In D128087#3599441, @Mordante wrote:

In D128087#3594642, @xingxue wrote:

Thanks very much, @Mordante and @philnik! There isn't an AIX maintainer IIUC but I will wait a couple of days to see if other reviewers have any comments. Thanks again!

You've added some of the people regularly working on AIX and @daltenty also works on AIX.

Add one more test case thousands_sep.pass.cpp.

Thanks very much, @Mordante and @philnik! There isn't an AIX maintainer IIUC but I will wait a couple of days to see if other reviewers have any comments. Thanks again!

Added missing "of" in the comment.

Added a comment in test case get_one.pass.cpp to explain the reason why the time field is omitted in the const definition for locale zh_CN.UTF-8.

Hi @ldionne, Would you please take a look and if no further comments, accept for #libc_abi? Thanks!

In D126915#3559129, @MaskRay wrote:

I see that https://discourse.llvm.org/t/rfc-adding-exported-visibility-style-to-the-ir-to-model-xcoff-exported-visibility/61853 says if no visibility bit (I assume n_type & SYM_V_MASK == 0) is set, the symbols need an export list.
I don't see how you use an export list for the libc++abi code, perhaps you maintain a down stream patch.

Thanks for your comments! Yes, export lists are currently used on AIX for libc++abi as well as other libraries downstream in the absence of attribute visibility.

Addressed comment - changed as per suggestion.

@MaskRay, @cebowleratibm, @ldionne, @hubert.reinterpretcast, @jasonliu: Thank you so much for your help, time and effort in reviewing, and constructive suggestions. All of these make the patch much better than it was. Thanks again!

Thanks very much, @MaskRay!
@ldionne Do you have further comments? If not, would you be able to approve for #libc_abi?

Hi @MaskRay, Do you have further comments? If not, would you be able to approve this patch? My colleague @cebowleratibm, who is an AIX compiler and runtime expert, has done a thorough review of AIX specific contents and approved it (see comment https://reviews.llvm.org/D100504#3414984). This implementation has been run through the proprietary test coverage and included in the IBM Open XL C/C++ for AIX V17.1 compiler released in Sept. 2021.

Deleted an empty line with a leading # sign as suggested.

Rebased and picked up the change in D125981 which introduces the cmake-bridge.cfg.in file to reduce the libunwind test config duplication. Thanks, @ldionne!

Addressed comments:

• Fixed formatting.

Hi @MaskRay, Do you have further comments?

Deleted a trailing empty line in file ibm-libunwind-shared.cfg.in.

Fixed problems in pre-merg check.

• Avoid using GCC extension named variadic macros;
• Avoid using C99 feature flexible array members.

Addressed comments. Moved the implementation for AIX state table based EH into a separate file aix_state_tab_eh.inc as suggested.

LGTM; however, will need the review from the libc++ group.

Addressed comments:

• Fixed typos in comments;
• Removed unnecessary cast.

@MaskRay @compnerd @cebowleratibm @sfertile Thanks so much for your time and effort in reviewing this patch and providing constructive suggestions to make it better! Much appreciated!

Gentle ping.

Gentle ping.

Fixed format according to git-clang-format.

Addressed comments.

• Updated code comments;
• Renamed variable inlineDestructor to cleanupLabel. Replaced the term "inline destructor" with "cleanup code";
• Renamed function compute_obj_addr to compute_addr_from_table.

Ping... This patch failed the pre-merge check because the formatting in lines 611-612 and 1179-1180 of file Registers.hpp follows the surrounding code rather than git-clang-format.

Addressed comments.

• Used unions where the same memory is used for different purposes to make overlapping use of the memory more clear.
• Renamed variables to be more descriptive.
• Added a high level description of the state table based EH.
• Added comments in various locations.
• Fixed a problem where the frame pointer is needed to calculate the address.
• Fixed the address calculation of virtual base object for calling destructor or deletor to be based on the IBM object model used by the proprietary xlC compiler because these paths are only used for xlC generated code.
• Commoned up some code to reduce redundancy.

Addressed comments:

• Remove the cursor size definition for building with AIX GCC for now because it is not enabled yet;
• Use the C++ style cast rather than C style cast;
• Be consistent with the surrounding area's formatting.

LGTM; thanks!

Addressed comments.

• fix errors in comments.

Addressed comments.

• allow the declaration of _Unwind_GetTextRelBase() on AIX.
• rename getFuncName to getFuncNameFromTBTable.
• rename clangFrame to frameWithEHInfo and rename classicFrame to frameWithXLEHStateTable.
• add comments to describe the code looking for eh_info and dealing with the complexity arising from some XL compiler versions use the wrong ordering of flag bits.

Addressed comment:

• Use the same logic for skipping the name_len and name fields.

Addressed comments:

• Make the declaration of _Unwind_GetDataRelBase cleaner in unwind.h.
• Rename enum member oldFrame to classicXLFrame.
• Use typedef of xlcxx_personality_v0_t in the declaration of __xlcxx_personality_v0().
• Rename variable stateOffset to stateTableOffset.
• Add commet to explain why dynamically resolving the personality for the state table.

Hi @compnerd, Thanks very much for your comments! I've made changes accordingly. Please let me know if you have further comments!

Thanks for reviewing this patch, @danielkiss, @MaskRay, @ldionne! Yes, target=powerpc-ibm-aix (32-bit) and target=powerpc64-ibm-aix (64-bit) are LIT features. Updated the patch as suggested. Thanks, @ldionne!