Skip to content

PTX: Fix regression in kernel metadata on LLVM 20. #811

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
maleadt merged 2 commits into from
May 27, 2026
Merged

PTX: Fix regression in kernel metadata on LLVM 20. #811

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
8c852f2
PTX: Fix regression in kernel metadata on LLVM 20.
maleadt May 27, 2026
7856743
Remove IR-level tests.
maleadt May 27, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
17 changes: 12 additions & 5 deletions src/ptx.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -186,11 +186,16 @@ function finish_module!(@nospecialize(job::CompilerJob{PTXCompilerTarget}),
# entry instead of re-reading the metadata.
annotations = Metadata[entry]

## kernel metadata
append!(annotations, [MDString("kernel"),
ConstantInt(Int32(1))])
# kernel metadata
#
# on LLVM >= 20 the `ptx_kernel` calling convention already marks the
# entry; the redundant "kernel" nvvm.annotation causes miscompilations.
if LLVM.version() < v"20"
append!(annotations, [MDString("kernel"),
ConstantInt(Int32(1))])
end

## expected CTA sizes
# expected CTA sizes
if job.config.target.minthreads !== nothing
bounds = ntuple(i -> i <= length(job.config.target.minthreads) ?
job.config.target.minthreads[i] : 1, 3)
Expand Down Expand Up @@ -234,7 +239,9 @@ function finish_module!(@nospecialize(job::CompilerJob{PTXCompilerTarget}),
end
end

push!(metadata(mod)["nvvm.annotations"], MDNode(annotations))
if length(annotations) > 1
push!(metadata(mod)["nvvm.annotations"], MDNode(annotations))
end
end

# we emit properties (of the device and ptx isa) as private global constants,
Expand Down
35 changes: 0 additions & 35 deletions test/ptx.jl
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -35,41 +35,6 @@ end
end
end

@testset "property_annotations" begin
@test @filecheck PTX.code_llvm(Tuple{}; dump_module=true) do
@check_not "nvvm.annotations"
return
end
@test @filecheck PTX.code_llvm(Tuple{}; dump_module=true, kernel=true) do
@check_not "maxntid"
@check_not "reqntid"
@check_not "minctasm"
@check_not "maxnreg"
@check "nvvm.annotations"
return
end
@test @filecheck PTX.code_llvm(Tuple{}; dump_module=true, kernel=true, maxthreads=42) do
@check "maxntidx\", i32 42"
@check "maxntidy\", i32 1"
@check "maxntidz\", i32 1"
return
end
@test @filecheck PTX.code_llvm(Tuple{}; dump_module=true, kernel=true, minthreads=42) do
@check "reqntidx\", i32 42"
@check "reqntidy\", i32 1"
@check "reqntidz\", i32 1"
return
end
@test @filecheck PTX.code_llvm(Tuple{}; dump_module=true, kernel=true, blocks_per_sm=42) do
@check "minctasm\", i32 42"
return
end
@test @filecheck PTX.code_llvm(Tuple{}; dump_module=true, kernel=true, maxregs=42) do
@check "maxnreg\", i32 42"
return
end
end

LLVM.version() >= v"8" && @testset "calling convention" begin
@test @filecheck PTX.code_llvm(Tuple{}; dump_module=true) do
@check_not "ptx_kernel"
Expand Down
Loading