That sounds fine to me; here is the original patch, with the updated test.

Address feedback

When I went to mark these as static I noticed they use CGDebugInfo::CreateMemberType which uses a couple other non-static member functions, and it starts to become difficult to tease things out into nice clean static functions.

Emit lifetime intrinsics for the sizes temp, and update test

I still don't quite see what you describe; with that change all of the lifetime.end calls pile up just before the enclosing function returns, not after each call to enqueue_kernel. Looking at https://clang.llvm.org/doxygen/EHScopeStack_8h_source.html#l00078 I don't see any option which isn't based on scope. The lifetime.start calls do occur where I would expect, though, so I will update the patch.

Update test

I have the library in BinaryFormat because it is used outside of the AMDGPU backend in tools (currently just llvm-readobj) in https://reviews.llvm.org/D48179

Address feedback

Address feedback; I hope I understood correctly what debug info to check for.

Add comments to explain OpenCL case

I noticed I had assumed AArch64 was available in the unittests, so I updated those to be skipped in the case where it isn't.

Sorry, I didn't see the additional comments until after I committed. I will make those changes; is it OK to update this review, or should I create a new one?

Rebase and ping

Rebase and ping

Rebase and ping

Rebase and ping

Thank you for taking a look @yaxunl. Should I wait for another reviewer or can I commit this?

Address feedback

I have updated the lit test with update_llc_test_checks, updated the condition in TargetLowering to allow for the case where DemandedElts is already wide enough, and added unittests for the cases which trigger the ICE in each function. This is still dependent on my assumptions about the exact semantics of EXTRACT_SUBVECTOR being correct.

I can update the test with update_llc_checks, but I'm not sure about testing the other functions. This test case is a minimized version of a test which happened to trigger the assert; to test the others would it be reasonable to use unit tests instead of LLVM IR?

Addressed feedback; updated other places where DemandedElts is zero extended to the source vector width.

Add tests for the the boundary condition for subregs.

Address some more feedback; use -stress-regalloc to cut down on the clobbers needed, add non-kernel tests, explicitly test the increment/decrement case, including the scratch offset SGPR.

Addressed feedback, and added at least one test to exercise the fix and the condition for putting the offset in an SGPR.

Ping; the use in AMDGPU's backend is now up on Phabricator.

Rebase

See https://reviews.llvm.org/D44429

Yes, I had misplaced 10000B.

Here is the simplified LIS in my case just before the assert. The failing SubRange is %8273.L00000001 which I believe is %8273.sub0

I think I am running into the same bug. I do not have a reduced test case yet, because it appears the bug is triggered due to spilling in my case.

LGTM

LGTM

Address feedback

Address feedback

Don't emit v2 kernel symbol type

Style fixes

Rebase, update syntax, expand tests

Address feedback

I missed reverting to the original AMDHSAKernelDescriptor.h

Do not output default values when printing .amdhsa_kernel directives