Preliminary Julia v1.11 support

.github/workflows/CI.yml

@@ -19,16 +19,16 @@ jobs:
       fail-fast: false
       matrix:
         version:
+          - 'pre'
           - '1.9'
-          - '~1.10.0-0'
-          - 'nightly'
+          - '1.10'
         os:
           - ubuntu-latest
         arch:
           - x64
     steps:
       - uses: actions/checkout@v3
-      - uses: julia-actions/setup-julia@v1
+      - uses: julia-actions/setup-julia@v2
         with:
           version: ${{ matrix.version }}
           arch: ${{ matrix.arch }}

src/AllocCheck.jl

@@ -204,8 +204,8 @@ function check_allocs(@nospecialize(func), @nospecialize(types); ignore_throw=tr
     target = DefaultCompilerTarget()
     job = CompilerJob(source, config)
     allocs = JuliaContext() do ctx
-        mod, meta = GPUCompiler.compile(:llvm, job, validate=false, optimize=false, cleanup=false)
-        optimize!(job, mod)
+        mod, meta = GPUCompiler.compile(:llvm, job, validate=false, optimize=true, cleanup=false)
+        # GPUCompiler.optimize!(job, mod) # TODO: Why is this broken?
 
         allocs = find_allocs!(mod, meta; ignore_throw)
         # display(mod)

src/classify.jl

@@ -62,7 +62,7 @@ function fn_may_allocate(name::AbstractString; ignore_throw::Bool)
     if name in ("egal__unboxed", "lock_value", "unlock_value", "get_nth_field_noalloc",
                 "load_and_lookup", "lazy_load_and_lookup", "box_bool", "box_int8",
                 "box_uint8", "excstack_state", "restore_excstack", "enter_handler",
-                "pop_handler", "f_typeof", "clock_now", "throw", "gc_queue_root", "gc_enable",
+                "pop_handler", "pop_handler_noexcept", "f_typeof", "clock_now", "throw", "gc_queue_root", "gc_enable",
                 "gc_disable_finalizers_internal", "gc_is_in_finalizer", "enable_gc_logging",
                 "gc_safepoint", "gc_collect", "genericmemory_owner", "get_pgcstack") || occursin(r"^unbox_.*", name)
         return false # these functions never allocate

src/compiler.jl

@@ -111,8 +111,8 @@ function compile_callable(f::F, tt::TT=Tuple{}; ignore_throw=true) where {F, TT}
 
         function compile(@nospecialize(job::CompilerJob))
             return JuliaContext() do ctx
-                mod, meta = GPUCompiler.compile(:llvm, job, validate=false)
-                optimize!(job, mod)
+                mod, meta = GPUCompiler.compile(:llvm, job, validate=false, optimize=true)
+                # optimize!(job, mod) # TODO
 
                 clone = copy(mod)
                 analysis = find_allocs!(mod, meta; ignore_throw)

src/compiler_utils.jl

@@ -32,7 +32,64 @@ function cpu_features()
     end
 end
 
-if VERSION >= v"1.10-beta3"
+mutable struct PipelineConfig
+    Speedup::Cint
+    Size::Cint
+    lower_intrinsics::Cint
+    dump_native::Cint
+    external_use::Cint
+    llvm_only::Cint
+    always_inline::Cint
+    enable_early_simplifications::Cint
+    enable_early_optimizations::Cint
+    enable_scalar_optimizations::Cint
+    enable_loop_optimizations::Cint
+    enable_vector_pipeline::Cint
+    remove_ni::Cint
+    cleanup::Cint
+end
+
+if VERSION >= v"1.11-alpha1"
+    function build_newpm_pipeline!(
+        pb::LLVM.PassBuilder,
+        mpm::LLVM.NewPMModulePassManager,
+        speedup=2,
+        size=0,
+        lower_intrinsics=true,
+        dump_native=false,
+        external_use=false,
+        llvm_only=false,
+        always_inline=true,
+        enable_early_simplifications=true,
+        enable_early_optimizations=true,
+        enable_scalar_optimizations=true,
+        enable_loop_optimizations=true,
+        enable_vector_pipeline=true,
+        remove_ni=true,
+        cleanup=false # note: modified vs. base
+    )
+        cfg = PipelineConfig(
+            speedup,
+            size,
+            lower_intrinsics,
+            dump_native,
+            external_use,
+            llvm_only,
+            always_inline,
+            enable_early_simplifications,
+            enable_early_optimizations,
+            enable_scalar_optimizations,
+            enable_loop_optimizations,
+            enable_vector_pipeline,
+            remove_ni,
+            cleanup,
+        )
+        ccall(:jl_build_newpm_pipeline, Cvoid,
+            (LLVM.API.LLVMModulePassManagerRef, LLVM.API.LLVMPassBuilderRef, Ref{PipelineConfig}),
+            mpm, pb, cfg
+        )
+    end
+elseif VERSION >= v"1.10-beta3"
     function build_newpm_pipeline!(pb::LLVM.PassBuilder, mpm::LLVM.NewPMModulePassManager, speedup=2, size=0, lower_intrinsics=true,
         dump_native=false, external_use=false, llvm_only=false,)
         ccall(:jl_build_newpm_pipeline, Cvoid,