From 8c852f26c2829da63d0e5eca830cef30cfd477ae Mon Sep 17 00:00:00 2001
From: Tim Besard <tim.besard@gmail.com>
Date: Wed, 27 May 2026 15:21:44 +0200
Subject: [PATCH 1/2] PTX: Fix regression in kernel metadata on LLVM 20.

---
 src/ptx.jl | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)

diff --git a/src/ptx.jl b/src/ptx.jl
index a3af659d..a7a01df6 100644
--- a/src/ptx.jl
+++ b/src/ptx.jl
@@ -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)
@@ -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,

From 7856743575631cbdc827416fae80bd9a856b5aac Mon Sep 17 00:00:00 2001
From: Tim Besard <tim.besard@gmail.com>
Date: Wed, 27 May 2026 17:08:27 +0200
Subject: [PATCH 2/2] Remove IR-level tests.

---
 test/ptx.jl | 35 -----------------------------------
 1 file changed, 35 deletions(-)

diff --git a/test/ptx.jl b/test/ptx.jl
index e1426ff2..f852f58a 100644
--- a/test/ptx.jl
+++ b/test/ptx.jl
@@ -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"