| author |  Robin Voetter <robin@voetter.nl> 2022-08-20 12:42:27+02:00 |
| committer | Robin Voetter <robin@voetter.nl> 2022-10-12 20:34:41+02:00 |
| log | 3392de87dcf516ed2e3fd1904372195bdc75c0eb |
| tree | b87b51283d6a78c708de7a9ee75b4acabfcf0d24 |
| parent | 5859d8458f919676be804835aaeb3ed844a67542 |
| signature |  Commit is signed but in an unrecognized format. |
allow global/local/shared address spaces on amdgcn
2 files changed, 12 insertions(+), 2 deletions(-)
src/Sema.zig+4-2
| ... | @@ -30303,13 +30303,15 @@ pub fn analyzeAddrspace( | ... | @@ -30303,13 +30303,15 @@ pub fn analyzeAddrspace( |
| 30303 | const address_space = addrspace_tv.val.toEnum(std.builtin.AddressSpace); | 30303 | const address_space = addrspace_tv.val.toEnum(std.builtin.AddressSpace); |
| 30304 | const target = sema.mod.getTarget(); | 30304 | const target = sema.mod.getTarget(); |
| 30305 | const arch = target.cpu.arch; | 30305 | const arch = target.cpu.arch; |
| 30306 | const is_gpu = arch == .nvptx or arch == .nvptx64; | 30306 | const is_nv = arch == .nvptx or arch == .nvptx64; |
| 30307 | const is_gpu = is_nv or arch == .amdgcn; | ||
| 30307 | 30308 | ||
| 30308 | const supported = switch (address_space) { | 30309 | const supported = switch (address_space) { |
| 30309 | .generic => true, | 30310 | .generic => true, |
| 30310 | .gs, .fs, .ss => (arch == .i386 or arch == .x86_64) and ctx == .pointer, | 30311 | .gs, .fs, .ss => (arch == .i386 or arch == .x86_64) and ctx == .pointer, |
| 30311 | // TODO: check that .shared and .local are left uninitialized | 30312 | // TODO: check that .shared and .local are left uninitialized |
| 30312 | .global, .param, .shared, .local => is_gpu, | 30313 | .param => is_nv, |
| 30314 | .global, .shared, .local => is_gpu, | ||
| 30313 | .constant => is_gpu and (ctx == .constant), | 30315 | .constant => is_gpu and (ctx == .constant), |
| 30314 | }; | 30316 | }; |
| 30315 | 30317 |
src/codegen/llvm.zig+8
| ... | @@ -2659,6 +2659,14 @@ pub const DeclGen = struct { | ... | @@ -2659,6 +2659,14 @@ pub const DeclGen = struct { |
| 2659 | .local => llvm.address_space.nvptx.local, | 2659 | .local => llvm.address_space.nvptx.local, |
| 2660 | else => unreachable, | 2660 | else => unreachable, |
| 2661 | }, | 2661 | }, |
| 2662 | .amdgcn => switch (address_space) { | ||
| 2663 | .generic => llvm.address_space.flat, | ||
| 2664 | .global => llvm.address_space.amdgpu.global, | ||
| 2665 | .constant => llvm.address_space.amdgpu.constant, | ||
| 2666 | .shared => llvm.address_space.amdgpu.local, | ||
| 2667 | .local => llvm.address_space.amdgpu.private, | ||
| 2668 | else => unreachable, | ||
| 2669 | }. | ||
| 2662 | else => switch (address_space) { | 2670 | else => switch (address_space) { |
| 2663 | .generic => llvm.address_space.default, | 2671 | .generic => llvm.address_space.default, |
| 2664 | else => unreachable, | 2672 | else => unreachable, |