Tracking Issue for amdgcn target

[Flakebi]

This issue tracks the implementation of the amdgpu target.

• MCP: Add amdgpu target compiler-team#823

Old, closed tracking issue: #51575

Current work-in-progress branch: https://github.com/Flakebi/rust/tree/amdgpu

Examples and support libraries: https://github.com/Flakebi/amdgpu-rs

Implementation

• Add amdgpu target: Add amdgpu target #134740
• Disable f128 in compiler-builtins: Disable f128 for amdgpu compiler-builtins#737
• Cast allocas to default address space: Cast allocas to default address space #135025
• Cast global vars to default address space: Cast global variables to default address space #135026
• Add triple to cc crate (once target is in nightly) (not necessary anymore)
• Add test compiling core and alloc
• Add gpu-kernel calling convention: Add gpu-kernel calling convention #135047
• Require no_mangle or export_name for gpu-kernel functions: Add checks for gpu-kernel calling conv #149991
• Export <kernel>.kd kernel descriptor symbol: Export kernel descriptor for amdgpu kernels #135909
• Require explicit -C target-cpu=: Target option to require explicit cpu #135030
• Fix linker-plugin-lto only doing thin lto: Fix linker-plugin-lto only doing thin lto #136840
• Fix requiring lto automatically: Fix requires_lto targets needing lto set in cargo #149624
• Remove range-metadata amdgpu workaround: Remove range-metadata amdgpu workaround #135027
• Add tests for the gpu-kernel calling convention lowering to amdgpu_kernel and the global variable addrspacecasts: Add amdgpu test for addrspacecasting global vars and the gpu-kernel calling convention #146148
• Support shared memory (dynamic): Tracking Issue for NVPTX shared memory #135516 Add intrinsic for launch-sized workgroup memory on GPUs #146181
• Support shared memory (static): Tracking Issue for NVPTX shared memory #135516
• Inline assembly: Add inline asm support for amdgpu #149793
• Allow vector types: Allow vector types for amdgpu #149994
• Get kernel dispatch packet pointer: Add amdgpu_dispatch_ptr intrinsic #150607
• Tracking Issue for amdgpu arch intrinsics: Tracking Issue for amdgpu arch intrinsics #149988

Bugs

• Debug compilation fails due to a bug in LLVM’s AMDGPUResourceUsageAnalysis. This is fixed in llvm main, so will be fixed with the update to LLVM 20
• rustc may incorrectly copy convergent operations: MIR passes do not take into account if an operation is convergent #137086 Do not run jump-threading for GPUs #149637 Use convergent attribute to funcs for GPU targets #153995
• Taking structs as arguments to gpu-kernel functions does not work correctly and should be forbidden: Add checks for gpu-kernel calling conv #149991
• Need -Clink-arg=--no-gc-sections when using linker-plugin-lto: AMDGPUTargetStreamer generates .kd symbols, breaking LTO requirement, may be discarded by --gc-sections llvm/llvm-project#119479
• MergeFunctions LLVM pass can generate invalid calls to kernels. Fixed in llvm main, so will be fixed with the next LLVM update: MergeFunctions LLVM pass can generate invalid function calls under calling convention #57356 [MergeFunc] Do not merge kernel functions llvm/llvm-project#174254