aboutsummaryrefslogtreecommitdiff
path: root/include/elf/amdgpu.h
AgeCommit message (Collapse)AuthorFilesLines
2022-03-16binutils/readelf: handle AMDGPU relocation typesSimon Marchi1-0/+21
Make readelf recognize AMDGPU relocation types, as documented here: https://llvm.org/docs/AMDGPUUsage.html#amdgpu-relocation-records The user-visible change looks like: -000000000004 000400000001 unrecognized: 1 0000000000000000 SCRATCH_RSRC_DWORD0 -00000000000c 000500000001 unrecognized: 1 0000000000000000 SCRATCH_RSRC_DWORD1 -000000000014 000600000007 unrecognized: 7 0000000000000000 global_var0 -00000000001c 000700000008 unrecognized: 8 0000000000000000 global_var1 -000000000024 000800000009 unrecognized: 9 0000000000000000 global_var2 -00000000002c 00090000000a unrecognized: a 0000000000000000 global_var3 -000000000034 000a0000000b unrecognized: b 0000000000000000 global_var4 +000000000004 000400000001 R_AMDGPU_ABS32_LO 0000000000000000 SCRATCH_RSRC_DWORD0 +00000000000c 000500000001 R_AMDGPU_ABS32_LO 0000000000000000 SCRATCH_RSRC_DWORD1 +000000000014 000600000007 R_AMDGPU_GOTPCREL 0000000000000000 global_var0 +00000000001c 000700000008 R_AMDGPU_GOTPCREL 0000000000000000 global_var1 +000000000024 000800000009 R_AMDGPU_GOTPCREL 0000000000000000 global_var2 +00000000002c 00090000000a R_AMDGPU_REL32_LO 0000000000000000 global_var3 +000000000034 000a0000000b R_AMDGPU_REL32_HI 0000000000000000 global_var4 binutils/ChangeLog: * readelf.c (dump_relocations): Handle EM_AMDGPU. include/ChangeLog: * elf/amdgpu.h: Add relocation values. Change-Id: I2ed4589f4cd37ea11ad2e0cb38d4b682271e1334
2022-03-16binutils/readelf: handle NT_AMDGPU_METADATA note nameSimon Marchi1-0/+4
Handle the NT_AMDGPU_METADATA note, which is described here: https://llvm.org/docs/AMDGPUUsage.html#code-object-v3-note-records As of this patch, just print out the name, not the contents, which is in the msgpack format. binutils/ChangeLog: * readelf.c (get_amdgpu_elf_note_type): New. (process_note): Handle "AMDGPU" notes. include/ChangeLog: * elf/amdgcn.h (NT_AMDGPU_METADATA): New. Change-Id: Id2dba2e2aeaa55ef7464fb35aee9c7d5f96ddb23
2022-03-16binutils/readelf: decode AMDGPU-specific e_flagsSimon Marchi1-0/+54
Decode and print the AMDGPU-specific fields of e_flags, as documented here: https://llvm.org/docs/AMDGPUUsage.html#header That is: - The specific GPU model - Whether the xnack and sramecc features are enabled The result looks like: - Flags: 0x52f + Flags: 0x52f, gfx906, xnack any, sramecc any The flags for the "HSA" OS ABI are properly versioned and documented on that page. But the NONE, PAL and MESA3D OS ABIs are not well documented nor versioned. Taking a peek at the LLVM source code, we see that they encode their flags the same way as HSA v3. For example, for PAL: https://github.com/llvm/llvm-project/blob/c8b614cd74a92d85936aed5ac7c642af75ffdc29/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp#L601 So for those other OS ABIs, we read them the same as HSA v3. binutils/ChangeLog: * readelf.c: Include elf/amdgcn.h. (decode_AMDGPU_machine_flags): New. (get_machine_flags): Handle flags for EM_AMDGPU machine type. include/ChangeLog: * elf/amdgcn.h: Add EF_AMDGPU_MACH_AMDGCN_* and EF_AMDGPU_FEATURE_* defines. Change-Id: Ib5b94df7cae0719a22cf4e4fd0629330e9485c12
2022-03-16bfd: add AMDGCN architectureSimon Marchi1-0/+36
Add support for the AMDGCN architecture to BFD. This is the bare minimum to get $ ./configure --target=amdgcn-hsa-amdhsa --disable-gas $ make all-binutils working later in this series. The specific AMDGCN models added here are a bit arbitrary, based on what we intend to initially support in GDB. This list will need to be updated in the future anyway. The complete up-to-date list of existing AMDGPU models can be found here: https://llvm.org/docs/AMDGPUUsage.html#processors The ELF format for this architecture is documented here: https://llvm.org/docs/AMDGPUUsage.html#elf-code-object The flags for the "HSA" OS ABI are properly versioned and documented on that page. But the NONE, PAL and MESA3D OS ABIs are not well documented nor versioned. Taking a peek at the LLVM source code, we see that they encode their flags the same way as HSA v3. For example, for PAL: https://github.com/llvm/llvm-project/blob/c8b614cd74a92d85936aed5ac7c642af75ffdc29/llvm/lib/Target/AMDGPU/MCTargetDesc/AMDGPUTargetStreamer.cpp#L601 So at least, we know that all AMDGPU objects (of which AMDGCN objects are a subset of) at the time of writing encode the specific GPU model in the EF_AMDGPU_MACH field of e_flags. bfd/ChangeLog: * Makefile.am (ALL_MACHINES, ALL_MACHINES_CFILES): Add cpu-amdgcn.c. (BFD64_BACKENDS): Add elf64-amdgcn.lo. (BFD64_BACKENDS_CFILES): Add elf64-amdgcn.c. * Makefile.in: Re-generate. * cpu-amdgcn.c: New. * elf64-amdgcn.c: New. * archures.c (bfd_architecture): Add bfd_arch_amdgcn and related mach defines. (bfd_amdgcn_arch): New. (bfd_archures_list): Add bfd_amdgcn_arch. * bfd-in2.h: Re-generate. * config.bfd: Handle amdgcn* target. * configure.ac: Handle amdgcn_elf64_le_vec. * configure: Re-generate. * elf-bfd.h (elf_target_id): Add AMDGCN_ELF_DATA. * targets.c (amdgcn_elf64_le_vec): New. (_bfd_target_vector): Add amdgcn_elf64_le_vec. include/ChangeLog: * elf/amdgpu.h: New. * elf/common.h (ELFOSABI_AMDGPU_HSA): Add. Change-Id: I969f7b14960797e88891c308749a6e341eece5b2
generated by cgit v1.1 at 2025-02-22 10:28:16 +0000