[luraess]

Since AMDGPU_LLVM_Backend_jll 23.1.1+3, the external back-end aligns i128 to 16 bytes. Julia before 1.12 aligns Int128 to 8 bytes on the host. Any kernel argument at or after an Int128 therefore ends up at a different offset on the device than on the host, which makes kernels fault or read garbage (JuliaGPU/AMDGPU.jl#1002). libamdgpu replaces the module's datalayout with its own. Since 23.1.1+4 (JuliaPackaging/Yggdrasil#15010), it keeps a 64-bit i128 alignment if the module asks for one.

For the external back-end, this PR therefore appends the host's i128 alignment to the GCN datalayout. On Julia 1.12+, that is i128:128. The kernel argument layout doesn't change there, but the middle end's datalayout now matches the back-end's instead of inheriting the 64-bit default of Julia's LLVM. The in-process back-end is unchanged, because its datalayout comes from Julia's LLVM.

The new test compares the kernel argument offsets and sizes in the amdhsa.kernels metadata with the host layout of the argument types. On Julia 1.11 it fails without the source change, and with JLL 23.1.1+3. It passes with this change and 23.1.1+4:

Julia 1.11, kernel(::Int32, ::Int128, ::Tuple{Int32,Int128}, ::Int64)	Int128 offset	tuple size
host	8	24
main, either JLL build	16	32
this PR, 23.1.1+3	16	32
this PR, 23.1.1+4	8	24

On 1.12 and 1.13 the layout matches the host with or without this PR (16 and 32).

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 88.43%. Comparing base (c8f5617) to head (948ac96).
⚠️ Report is 8 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #983   +/-   ##
=======================================
  Coverage   88.42%   88.43%           
=======================================
  Files          30       30           
  Lines        6395     6400    +5     
=======================================
+ Hits         5655     5660    +5     
  Misses        740      740           

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[luraess]

CI was all green, I just now trimmed the inline comments. if it looks good, happy to have it merged 🙏

[simeonschaub]

Changes look good, I wonder whether we can add an @filecheck regression test to check the backend actually does the right thing, i.e. something that fails with the current AMDGPU_LLVM_Backend_jll, but won't after JuliaPackaging/Yggdrasil#15010

[luraess]

I wonder whether we can add an @filecheck regression test

Good point. I'll see to add that

[maleadt]

I just now trimmed the inline comments

Better use [ci skip] then to avoid wasting CI resources.

[luraess]

Ah yeah right

[maleadt]

Added a test.

[luraess]

Can this be merged and a released so we can finalise JuliaGPU/AMDGPU.jl#1139 🙏

[luraess]

Thanks! Would it be possible to tag and release a patch?