v0.26.5

GPUCompiler v0.26.5

Diff since v0.26.4

Merged pull requests:

• Update New PassManager Pipeline (#559) (@vchuravy)
• Add missing backticks (#566) (@jonas-schulze)
• Add a testsuite for inference caching (#567) (@vchuravy)
• Fixes for 1.12 (#568) (@maleadt)
• Don't opt into typeinf tracking before v1.11 (#569) (@vchuravy)
• [Metal] Add check for i128 usage (#571) (@tgymnich)
• Add an interface extension point to prepare before anything happens. (#572) (@maleadt)
• Make the optimization level configurable. (#577) (@maleadt)
• Bump julia-actions/setup-julia from 1 to 2 (#578) (@dependabot[bot])
• Adapt to LLVM 17. (#583) (@maleadt)
• Fixes for LLVM 17 (#584) (@maleadt)

Closed issues:

• PTX: Don't lower unreachable on LLVM 17 (#475)
• Opaque pointers support (#511)
• Readiness for LLVM 17 (#573)

v0.26.4

GPUCompiler v0.26.4

Diff since v0.26.3

Merged pull requests:

• Metal: Work around function attribute changes in LLVM 16. (#564) (@maleadt)

v0.26.3

GPUCompiler v0.26.3

Diff since v0.26.2

Merged pull requests:

• Adapt to upcoming Julia changes (#562) (@maleadt)
• Include the compiler config in the runtime cache lookup. (#563) (@maleadt)

Closed issues:

• Cache ignores compiler target properties (#560)

v0.26.2

GPUCompiler v0.26.2

Diff since v0.26.1

Merged pull requests:

• Support internal cache (#545) (@vchuravy)
• Use the new jl_method_lookup_by_tt API on 1.11. (#550) (@maleadt)
• Add CI for 1.11. (#551) (@maleadt)
• Specialize MI if necessary (#553) (@vchuravy)
• Fix construction of byval attribute. (#555) (@maleadt)

Closed issues:

• methodinstance(foo, types) returns MethodError permanently if method is not defined (#530)
• SPIRV_LLVM_Translator needs build for LLVM 16/17 (#543)
• Metal calling convention in LLVM used by ARM in LLVM 16 (#544)

v0.26.1

GPUCompiler v0.26.1

Diff since v0.26.0

Merged pull requests:

• Metal: Run ExpandReductions pass. (#549) (@maleadt)

v0.26.0

GPUCompiler v0.26.0

Diff since v0.25.0

Merged pull requests:

• Bump actions/checkout from 3 to 4 (#525) (@dependabot[bot])
• Update manifest (#526) (@github-actions[bot])
• Support and test opaque pointers (#528) (@maleadt)
• use CC.add_invalidation_callback! if available (#529) (@aviatesk)
• Fix CodeCache() constructor. (#533) (@vchuravy)
• Fix and bound Aqua tests. (#537) (@maleadt)
• Fix type suffix of integer intrinsics (#538) (@tgymnich)
• Bump codecov/codecov-action from 3 to 4 (#539) (@dependabot[bot])
• Bump actions/cache from 3 to 4 (#540) (@dependabot[bot])
• Adjust for upstream JuliaLang/julia#53088 (#541) (@vchuravy)
• Skip SPIRV tests if tooling is not available (#542) (@vchuravy)
• Metal: Remove ability to generate binary libraries. (#548) (@maleadt)

Closed issues:

• Overlay methods disabling IR interpreter breaks const-prop of GPU-incompatible code (#384)
• GCN: Assertion failure due to stack address space (#504)
• Optimizer regression on 1.11 (#506)
• propagate_julia_addrsp! is not defined (#531)

v0.25.0

GPUCompiler v0.25.0

Breaking change: PTX back-end now does not replace trap with exit anymore. This probably only matters to CUDA.jl, which now has to make sure to emit exit in its signal handler, to avoid tripping over the trap that probably follows it.

Diff since v0.24.5

Merged pull requests:

• Update manifest (#522) (@github-actions[bot])
• PTX: Improve handling of trap (#524) (@maleadt)

v0.24.5

GPUCompiler v0.24.5

Diff since v0.24.4

Merged pull requests:

• Only warn about missed transformations when debug output is enabled. (#521) (@maleadt)

v0.24.4

GPUCompiler v0.24.4

Diff since v0.24.3

Merged pull requests:

• Metal: Emit name metadata. (#518) (@maleadt)

v0.24.3

GPUCompiler v0.24.3

Diff since v0.24.2

Merged pull requests:

• Metal: Support vector-valued intrinsics. (#515) (@maleadt)
• Disable the vectorizer pipeline. (#516) (@maleadt)
• Metal: Avoid duplicate intrinsics. (#517) (@maleadt)