One of our PowerPC buildbots is failing because it is named lld-multistage, which matches with this CHECK-NOT in clang/test/Driver/hip-sanitize-options.hip: ;CHECK-NOT: {{"[^"]*lld[^"]*".* ".*hip.bc"}}.
I've created a patch and if you could review it that'd be great!

In D97336#2585442, @phosek wrote:

In D97336#2585434, @Conanap wrote:

If this patch could be pushed soon that'd be great; we need it to unblock PowerPC buildbots.

I've reverted the original change D97110, are you still seeing the issue?

If this patch could be pushed soon that'd be great; we need it to unblock PowerPC buildbots.

Addressed some nits and a problem where sometimes the compiler would spill half way through materialization.

In D97110#2583184, @phosek wrote:

In D97110#2583082, @phosek wrote:

In D97110#2583026, @Conanap wrote:

Hello,

Power PC is also experiencing the same error,

Malformed instrumentation profile data
 error: No profiles could be merged.

I built a branch of main with this change reverted and the error goes away. Would it be possible to have this reverted for now until a proper solution has been created? Thanks!

I've managed to reproduce this locally in a Docker container. It's a bug in the older binutils bfd ld, the __start_ and __stop_ have incorrect addresses, it's working correctly in gold and also in newer binutils versions. I'll send a change to conditionally enable this feature with -fbinutils-version=, I'm just trying to identify what version of binutils this was fixed in.

I have bisected this to binutils 2.28 and I have sent D97336 to conditionalize the use of nobits to binutils >=2.28.

Power PC is also experiencing the same error,

Malformed instrumentation profile data
 error: No profiles could be merged.

Fixed up a few comments

Updated to ensure the shortcircuit protects against the destructive function.

Added a note on one of the functions

Addressed comments, fixed an incorrect encoding.

Removed lambda functions and made a static function, included a R2SaveStub test case that covers if it fits in 34 bits with option --no-power10-stubs.

Addressed comments

Fixed some formatting and comments.

Pushed; differential revision link accidentally had an extra https:// so it did not automatically close.

Loads Hi before Lo now; removed implicit cast.

Fixed a typo that prevented successful builds

Removed unecessary ternary

initialize as a form of undef instead

Addressed some comments for formatting and style

Updated tests as per Amy and Lei's comments

Will take over patch from Kamau

Removed braces for single lines

addressed a comment

Replied to a comment

Addressed some formatting comments

clang format

Clang-format

Updated a test case

Updated a test file.

The unrelated change no longer shows as diff, thank you!

Addressed some formatting problems as well as corrected incorrect arguments for Hi case.

Addressed Nemenja's concerns.

Use regex match to check for variable order in IR instead.

Addressed comments and corrected incorrect behaviour on Power10

Changed the location of fix and added extra test cases

Committed with Nemanja's comments addressed in the commit. Hash d7eb917a7cb793f49e16841fc24826b988dd5c8f

It looks like this commit is causing a few failures on nearly all PPC bots and a sanitizer bot; would it be possible to revert this commit for now until the issue is resolved?

Changed implementation for vrlqnm as per Nemanja

lgtm

lgtm

Open Coded instead

Open coded VMODS/UQ instead.

Added extra test case to demonstrate that the arguments will be reversed

LGTM

Looks like there's some unit test failures; could you double check?

Added --no-show-raw-insn and updated test case accordingly

Added back the test files that were accidentally removed in prev diff

Added an extra test case

Minor nit, okay if changed for commit

Added test case for non-zero offset, added a few flags to the RUN lines and removed some unecessary comments.

Taking over for the patch

Included extra test cases

Added shl tests, formatting fixes