Merge pull request #631 from JuliaGPU/tb/nightly

Re-enable nightly testing.

.github/workflows/ci.yml

@@ -18,12 +18,13 @@ jobs:
     strategy:
       fail-fast: false
       matrix:
-        version: ['1.10', 'pre'] # 'nightly'
+        version: ['1.10', 'pre', 'nightly']
         os: [ubuntu-latest, macOS-latest, windows-latest]
         arch: [x64]
         llvm_args: ['']
         include:
           # starting with Julia 1.10, we can enable opaque pointers
+          # from Juila 1.12 on, this is the default.
           - version: '1.10'
             os: 'ubuntu-latest'
             arch: 'x64'
@@ -48,18 +49,6 @@ jobs:
             os: 'windows-latest'
             arch: 'x64'
             llvm_args: '--opaque-pointers'
-          #- version: 'nightly'
-          #  os: 'ubuntu-latest'
-          #  arch: 'x64'
-          #  llvm_args: '--opaque-pointers'
-          #- version: 'nightly'
-          #  os: 'macOS-latest'
-          #  arch: 'x64'
-          #  llvm_args: '--opaque-pointers'
-          #- version: 'nightly'
-          #  os: 'windows-latest'
-          #  arch: 'x64'
-          #  llvm_args: '--opaque-pointers'
     steps:
       - uses: actions/checkout@v4
 

src/interface.jl

@@ -252,7 +252,11 @@ method_table(@nospecialize(job::CompilerJob)) = GLOBAL_METHOD_TABLE
 
 # the inference parameters to use when constructing the GPUInterpreter
 function inference_params(@nospecialize(job::CompilerJob))
-    return CC.InferenceParams(; unoptimize_throw_blocks=false)
+    if VERSION >= v"1.12.0-DEV.1017"
+        CC.InferenceParams()
+    else
+        CC.InferenceParams(; unoptimize_throw_blocks=false)
+    end
 end
 
 # the optimization parameters to use when constructing the GPUInterpreter

src/utils.jl

@@ -95,6 +95,12 @@ end
 # XXX: it's not allowed to switch tasks while under this lock, can we guarantee that?
 #      its probably easier to start using our own LLVM context when that's possible.
 macro locked(ex)
+    if VERSION >= v"1.12.0-DEV.769"
+        # no need to handle locking; it's taken care of by the engine
+        # as long as we use a correct cache owner token.
+        return esc(ex)
+    end
+
     def = splitdef(ex)
     def[:body] = quote
         ccall(:jl_typeinf_lock_begin, Cvoid, ())
@@ -109,6 +115,10 @@ end
 
 # HACK: temporarily unlock again to perform a task switch
 macro unlocked(ex)
+    if VERSION >= v"1.12.0-DEV.769"
+        return esc(ex)
+    end
+
     def = splitdef(ex)
     def[:body] = quote
         ccall(:jl_typeinf_lock_end, Cvoid, ())

src/validation.jl

@@ -218,6 +218,18 @@ function check_ir!(job, errors::Vector{IRError}, inst::LLVM.CallInst)
                 @safe_debug "Decoding arguments to jl_get_binding_or_error failed" inst bb=LLVM.parent(inst)
                 push!(errors, (DELAYED_BINDING, bt, nothing))
             end
+        elseif fn == "jl_reresolve_binding_value_seqcst" || fn == "ijl_reresolve_binding_value_seqcst"
+            try
+                # pry the binding from the IR
+                expr = arguments(inst)[1]::ConstantExpr
+                expr = first(operands(expr))::ConstantInt # get rid of inttoptr
+                ptr = Ptr{Any}(convert(Int, expr))
+                obj = Base.unsafe_pointer_to_objref(ptr)
+                push!(errors, (DELAYED_BINDING, bt, obj.globalref))
+            catch e
+                @safe_debug "Decoding arguments to jl_reresolve_binding_value_seqcst failed" inst bb=LLVM.parent(inst)
+                push!(errors, (DELAYED_BINDING, bt, nothing))
+            end
         elseif fn == "jl_invoke" || fn == "ijl_invoke"
             try
                 f, args, nargs, meth = arguments(inst)

test/gcn_tests.jl

@@ -74,44 +74,51 @@ end
     # bug: depending on a child function from multiple parents resulted in
     #      the child only being present once
 
-    @noinline child(i) = sink_gcn(i)
-    function parent1(i)
-        child(i)
-        return
+    mod = @eval module $(gensym())
+        export child, parent1, parent2
+
+        @noinline child(i) = sink_gcn(i)
+        function parent1(i)
+            child(i)
+            return
+        end
+        function parent2(i)
+            child(i+1)
+            return
+        end
     end
 
-    asm = sprint(io->GCN.code_native(io, parent1, Tuple{Int}; dump_module=true))
+    asm = sprint(io->GCN.code_native(io, mod.parent1, Tuple{Int}; dump_module=true))
     @test occursin(r"\.type.*julia_[[:alnum:]_.]*child_\d*,@function", asm)
 
-    function parent2(i)
-        child(i+1)
-        return
-    end
-
-    asm = sprint(io->GCN.code_native(io, parent2, Tuple{Int}; dump_module=true))
+    asm = sprint(io->GCN.code_native(io, mod.parent2, Tuple{Int}; dump_module=true))
     @test occursin(r"\.type.*julia_[[:alnum:]_.]*child_\d*,@function", asm)
 end
 
 @testset "child function reuse bis" begin
     # bug: similar, but slightly different issue as above
     #      in the case of two child functions
-    @noinline child1(i) = sink_gcn(i)
-    @noinline child2(i) = sink_gcn(i+1)
-    function parent1(i)
-        child1(i) + child2(i)
-        return
+
+    mod = @eval module $(gensym())
+        export parent1, parent2, child1, child2
+
+        @noinline child1(i) = sink_gcn(i)
+        @noinline child2(i) = sink_gcn(i+1)
+        function parent1(i)
+            child1(i) + child2(i)
+            return
+        end
+        function parent2(i)
+            child1(i+1) + child2(i+1)
+            return
+        end
     end
 
-    asm = sprint(io->GCN.code_native(io, parent1, Tuple{Int}; dump_module=true))
+    asm = sprint(io->GCN.code_native(io, mod.parent1, Tuple{Int}; dump_module=true))
     @test occursin(r"\.type.*julia_[[:alnum:]_.]*child1_\d*,@function", asm)
     @test occursin(r"\.type.*julia_[[:alnum:]_.]*child2_\d*,@function", asm)
 
-    function parent2(i)
-        child1(i+1) + child2(i+1)
-        return
-    end
-
-    asm = sprint(io->GCN.code_native(io, parent2, Tuple{Int}; dump_module=true))
+    asm = sprint(io->GCN.code_native(io, mod.parent2, Tuple{Int}; dump_module=true))
     @test occursin(r"\.type.*julia_[[:alnum:]_.]*child1_\d*,@function", asm)
     @test occursin(r"\.type.*julia_[[:alnum:]_.]*child2_\d*,@function", asm)
 end

test/native_tests.jl

@@ -388,46 +388,61 @@ Base.unsafe_trunc(::Type{Int}, x::CleverType) = unsafe_trunc(Int, x.x)
 end
 
 @testset "invalid LLVM IR" begin
-    foobar(i) = println(i)
+    mod = @eval module $(gensym())
+        export foobar
+        foobar(i) = println(i)
+    end
 
     @test_throws_message(InvalidIRError,
-                         Native.code_execution(foobar, Tuple{Int})) do msg
+                         Native.code_execution(mod.foobar, Tuple{Int})) do msg
         occursin("invalid LLVM IR", msg) &&
         (occursin(GPUCompiler.RUNTIME_FUNCTION, msg) ||
          occursin(GPUCompiler.UNKNOWN_FUNCTION, msg) ||
          occursin(GPUCompiler.DYNAMIC_CALL, msg)) &&
         occursin("[1] println", msg) &&
-        occursin(r"\[2\] .*foobar", msg)
+        occursin("[2] foobar", msg)
     end
 end
 
 @testset "invalid LLVM IR (ccall)" begin
-    foobar(p) = (unsafe_store!(p, ccall(:time, Cint, ())); nothing)
+    mod = @eval module $(gensym())
+        export foobar
+        function foobar(p)
+            unsafe_store!(p, ccall(:time, Cint, ()))
+            return
+        end
+    end
 
     @test_throws_message(InvalidIRError,
-                         Native.code_execution(foobar, Tuple{Ptr{Int}})) do msg
+                         Native.code_execution(mod.foobar, Tuple{Ptr{Int}})) do msg
         if VERSION >= v"1.11-"
             occursin("invalid LLVM IR", msg) &&
             occursin(GPUCompiler.LAZY_FUNCTION, msg) &&
             occursin("call to time", msg) &&
-            occursin(r"\[1\] .*foobar", msg)
+            occursin("[1] foobar", msg)
         else
             occursin("invalid LLVM IR", msg) &&
             occursin(GPUCompiler.POINTER_FUNCTION, msg) &&
-            occursin(r"\[1\] .*foobar", msg)
+            occursin("[1] foobar", msg)
         end
     end
 end
 
 @testset "delayed bindings" begin
-    kernel() = (undefined; return)
+    mod = @eval module $(gensym())
+        export kernel
+        function kernel()
+            undefined
+            return
+        end
+    end
 
     @test_throws_message(InvalidIRError,
-                         Native.code_execution(kernel, Tuple{})) do msg
+                         Native.code_execution(mod.kernel, Tuple{})) do msg
         occursin("invalid LLVM IR", msg) &&
         occursin(GPUCompiler.DELAYED_BINDING, msg) &&
-        occursin("use of 'undefined'", msg) &&
-        occursin(r"\[1\] .*kernel", msg)
+        occursin(r"use of '.*undefined'", msg) &&
+        occursin("[1] kernel", msg)
     end
 end
 
@@ -442,15 +457,18 @@ end
         occursin("invalid LLVM IR", msg) &&
         occursin(GPUCompiler.DYNAMIC_CALL, msg) &&
         occursin("call to nospecialize_child", msg) &&
-        occursin(r"\[1\] kernel", msg)
+        occursin("[1] kernel", msg)
     end
 end
 
 @testset "dynamic call (apply)" begin
-    func() = println(1)
+    mod = @eval module $(gensym())
+        export func
+        func() = println(1)
+    end
 
     @test_throws_message(InvalidIRError,
-                         Native.code_execution(func, Tuple{})) do msg
+                         Native.code_execution(mod.func, Tuple{})) do msg
         occursin("invalid LLVM IR", msg) &&
         occursin(GPUCompiler.DYNAMIC_CALL, msg) &&
         occursin("call to println", msg) &&