Use LLVM's byval instead of rewriting kernels.

[maleadt]

No description provided.

[codecov]

Codecov Report

Merging #16 into master will decrease coverage by 1.43%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master      #16      +/-   ##
==========================================
- Coverage   68.16%   66.73%   -1.44%     
==========================================
  Files          18       18              
  Lines        1087     1034      -53     
==========================================
- Hits          741      690      -51     
+ Misses        346      344       -2     

Impacted Files	Coverage Δ
src/irgen.jl	72.19% <100.00%> (-4.34%)	⬇️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 7fd5e2e...398b8ac. Read the comment docs.

[maleadt]

@jpsamaroo this triggers a bunch of address space-related issues with AMDGPUnative... Are those related to #13? Because the GCN tests in GPUCompiler do pass.

[vchuravy]

No those are new https://github.com/llvm/llvm-project/blob/d359f2096850c68b708bc25a7baca4282945949f/llvm/lib/Target/AMDGPU/AMDGPUAliasAnalysis.cpp#L91-L100

By rewriting the kernel we have sofar tricked out the AA for AMDGPU.

[maleadt]

Ah, this only used to fail when testing AMDGPUnative, not during GPUCompiler's GCN tests (maybe codegen tests covering the same functionality should be put here?), and AMD CI is down...

[jpsamaroo]

AMDGPU CI should be back up. The issues with "pointer address space out of range" are due to Julia's optimization pipeline running the verifier at the beginning on debug builds, which #30 should hopefully fix for GCN.

[maleadt]

Turns out there's a problem with this on Julia 1.4: NVPTX lowers byval by creating an alloca in the stack AS, which conflicts when passing a pointer in another address space: https://github.com/llvm-mirror/llvm/blob/2c4ca6832fa6b306ee6a7010bfb80a3f2596f824/lib/Target/NVPTX/NVPTXLowerArgs.cpp#L141-L168
Not a problem on 1.5, where we strip address spaces.