v1.23.0

GPUCompiler v1.23.0

Diff since v1.22.7

Merged pull requests:

• GCN: generate machine code via external AMDGPU_LLVM_Backend_jll (#857) (@vchuravy)
• Add alloca intrinsic for per-workitem stack scratch (#859) (@vchuravy)
• Turn Julia passes into closures (#864) (@vchuravy)
• Cover the function definition line of device code. (#866) (@maleadt)
• Fixes for nightly (#868) (@maleadt)
• Make the runtime library disk cache best-effort. (#869) (@maleadt)
• Fixes for nightly (#870) (@maleadt)

v1.22.7

GPUCompiler v1.22.7

Diff since v1.22.6

Merged pull requests:

• Metal: fix byte-GEP coalescing crash on cross-block chains (#863) (@maleadt)

v1.22.6

GPUCompiler v1.22.6

Diff since v1.22.5

Merged pull requests:

• Bump actions/checkout from 6 to 7 (#860) (@dependabot[bot])
• Remove temporary per-compile IR dump that taxed all downstream CI (#861) (@maleadt)
• Metal: work around a miscompile. (#862) (@maleadt)

v1.22.5

GPUCompiler v1.22.5

Diff since v1.22.4

Merged pull requests:

• TTFX-related improvements (#856) (@maleadt)

v1.22.4

GPUCompiler v1.22.4

Diff since v1.22.3

Merged pull requests:

• Restore current_job across nested compilation. (#855) (@maleadt)

Closed issues:

• Crash during kernel state pass (#851)

v1.22.3

GPUCompiler v1.22.3

Diff since v1.22.2

Merged pull requests:

• Harden compiler-error display against disposed LLVM modules (#854) (@maleadt)

v1.22.2

GPUCompiler v1.22.2

Diff since v1.22.1

Merged pull requests:

• Align runtime datalayout when relinking after non-integral stripping (#852) (@maleadt)
• Metal: mark by-reference kernel arguments read-only (#853) (@maleadt)

v1.22.1

GPUCompiler v1.22.1

Diff since v1.22.0

Merged pull requests:

• Fix handling of Julia's non IPO-safe metadata (#850) (@maleadt)

v1.22.0

GPUCompiler v1.22.0

Diff since v1.21.0

Merged pull requests:

• [Metal] Add quadgroup attributes (#835) (@christiangnrd)
• Gate aggressive instcombine pass (#842) (@wsmoses)
• Record line coverage of device code at compile time. (#844) (@maleadt)
• Require LLVM 9.9 and use the new debug location accessors. (#845) (@maleadt)
• Update Buildkite pipeline for the new JuliaGPU cluster (#846) (@maleadt)
• Move Metal.jl testing back to the main pipeline (#847) (@maleadt)
• Metal: Fixes for dynamic allocations (#848) (@maleadt)
• Metal: Add constant global address space hook (#849) (@maleadt)

v1.21.0

GPUCompiler v1.21.0

Diff since v1.20.0

Merged pull requests:

• Metal: Add a pass to convert software-emulated BFloat16 to the correct type. (#843) (@maleadt)