Default to using NewPM. (#604)

Author: maleadt

src/GPUCompiler.jl

@@ -16,8 +16,6 @@ using Preferences
 const CC = Core.Compiler
 using Core: MethodInstance, CodeInstance, CodeInfo
 
-const use_newpm = LLVM.has_newpm()
-
 include("utils.jl")
 include("mangling.jl")
 

src/driver.jl

@@ -291,8 +291,8 @@ const __llvm_initialized = Ref(false)
                     push!(preserved_gvs, LLVM.name(gvar))
                 end
             end
-            if use_newpm && LLVM.version() >= v"17"
-                run!(InternalizePass(InternalizePassOptions(; preserved_gvs)), ir,
+            if LLVM.version() >= v"17"
+                run!(InternalizePass(; preserved_gvs), ir,
                      llvm_machine(job.config.target))
             else
                 @dispose pm=ModulePassManager() begin
@@ -318,35 +318,17 @@ const __llvm_initialized = Ref(false)
                 # which also need to happen _after_ regular optimization.
                 # XXX: make these part of the optimizer pipeline?
                 if has_deferred_jobs
-                    if use_newpm
-                        @dispose pb=NewPMPassBuilder() begin
-                            add!(pb, NewPMFunctionPassManager()) do fpm
-                                add!(fpm, InstCombinePass())
-                            end
-                            add!(pb, AlwaysInlinerPass())
-                            add!(pb, NewPMFunctionPassManager()) do fpm
-                                add!(fpm, SROAPass())
-                                add!(fpm, GVNPass())
-                            end
-                            add!(pb, MergeFunctionsPass())
-                            run!(pb, ir, llvm_machine(job.config.target))
+                    @dispose pb=NewPMPassBuilder() begin
+                        add!(pb, NewPMFunctionPassManager()) do fpm
+                            add!(fpm, InstCombinePass())
                         end
-                    else
-                        @dispose pm=ModulePassManager() begin
-                            # inline and optimize the call to e deferred code. in particular we want
-                            # to remove unnecessary alloca's created by pass-by-ref semantics.
-                            instruction_combining!(pm)
-                            always_inliner!(pm)
-                            scalar_repl_aggregates_ssa!(pm)
-                            promote_memory_to_register!(pm)
-                            gvn!(pm)
-
-                            # merge duplicate functions, since each compilation invocation emits everything
-                            # XXX: ideally we want to avoid emitting these in the first place
-                            merge_functions!(pm)
-
-                            run!(pm, ir)
+                        add!(pb, AlwaysInlinerPass())
+                        add!(pb, NewPMFunctionPassManager()) do fpm
+                            add!(fpm, SROAPass())
+                            add!(fpm, GVNPass())
                         end
+                        add!(pb, MergeFunctionsPass())
+                        run!(pb, ir, llvm_machine(job.config.target))
                     end
                 end
             end
@@ -357,29 +339,13 @@ const __llvm_initialized = Ref(false)
 
         if cleanup
             @timeit_debug to "clean-up" begin
-                if use_newpm
-                    @dispose pb=NewPMPassBuilder() begin
-                        add!(pb, RecomputeGlobalsAAPass())
-                        add!(pb, GlobalOptPass())
-                        add!(pb, GlobalDCEPass())
-                        add!(pb, StripDeadPrototypesPass())
-                        add!(pb, ConstantMergePass())
-                        run!(pb, ir, llvm_machine(job.config.target))
-                    end
-                else
-                    # we can only clean-up now, as optimization may lower or introduce calls to
-                    # functions from the GPU runtime (e.g. julia.gc_alloc_obj -> gpu_gc_pool_alloc)
-                    @dispose pm=ModulePassManager() begin
-                        # eliminate all unused internal functions
-                        global_optimizer!(pm)
-                        global_dce!(pm)
-                        strip_dead_prototypes!(pm)
-
-                        # merge constants (such as exception messages)
-                        constant_merge!(pm)
-
-                        run!(pm, ir)
-                    end
+                @dispose pb=NewPMPassBuilder() begin
+                    add!(pb, RecomputeGlobalsAAPass())
+                    add!(pb, GlobalOptPass())
+                    add!(pb, GlobalDCEPass())
+                    add!(pb, StripDeadPrototypesPass())
+                    add!(pb, ConstantMergePass())
+                    run!(pb, ir, llvm_machine(job.config.target))
                 end
             end
         end

src/irgen.jl

@@ -101,9 +101,9 @@ function irgen(@nospecialize(job::CompilerJob))
         for gvar in globals(mod)
             push!(preserved_gvs, LLVM.name(gvar))
         end
-        if use_newpm && LLVM.version() >= v"17"
+        if LLVM.version() >= v"17"
             @dispose pb=NewPMPassBuilder() begin
-                add!(pb, InternalizePass(InternalizePassOptions(; preserved_gvs)))
+                add!(pb, InternalizePass(; preserved_gvs))
                 add!(pb, AlwaysInlinerPass())
                 run!(pb, mod, llvm_machine(job.config.target))
             end
@@ -718,11 +718,7 @@ function add_kernel_state!(mod::LLVM.Module)
 
     return true
 end
-add_kernel_state!(pm::PassManager) =
-    add!(pm, ModulePass("AddKernelStatePass", add_kernel_state!))
-if LLVM.has_newpm()
-    AddKernelStatePass() = NewPMModulePass("AddKernelStatePass", add_kernel_state!)
-end
+AddKernelStatePass() = NewPMModulePass("AddKernelStatePass", add_kernel_state!)
 
 # lower calls to the state getter intrinsic. this is a two-step process, so that the state
 # argument can be added before optimization, and that optimization can introduce new uses
@@ -774,11 +770,7 @@ function lower_kernel_state!(fun::LLVM.Function)
 
     return changed
 end
-lower_kernel_state!(pm::PassManager) =
-    add!(pm, FunctionPass("LowerKernelStatePass", lower_kernel_state!))
-if LLVM.has_newpm()
-    LowerKernelStatePass() = NewPMFunctionPass("LowerKernelStatePass", lower_kernel_state!)
-end
+LowerKernelStatePass() = NewPMFunctionPass("LowerKernelStatePass", lower_kernel_state!)
 
 function cleanup_kernel_state!(mod::LLVM.Module)
     job = current_job::CompilerJob
@@ -796,11 +788,7 @@ function cleanup_kernel_state!(mod::LLVM.Module)
 
     return changed
 end
-cleanup_kernel_state!(pm::PassManager) =
-    add!(pm, ModulePass("CleanupKernelStatePass", cleanup_kernel_state!))
-if LLVM.has_newpm()
-    CleanupKernelStatePass() = NewPMModulePass("CleanupKernelStatePass", cleanup_kernel_state!)
-end
+CleanupKernelStatePass() = NewPMModulePass("CleanupKernelStatePass", cleanup_kernel_state!)
 
 function kernel_state_intr(mod::LLVM.Module, T_state)
     state_intr = if haskey(functions(mod), "julia.gpu.state_getter")

src/mcgen.jl

@@ -6,29 +6,16 @@ function prepare_execution!(@nospecialize(job::CompilerJob), mod::LLVM.Module)
     global current_job
     current_job = job
 
-    if use_newpm
-        @dispose pb=NewPMPassBuilder() begin
-            register!(pb, ResolveCPUReferencesPass())
+    @dispose pb=NewPMPassBuilder() begin
+        register!(pb, ResolveCPUReferencesPass())
 
-            add!(pb, RecomputeGlobalsAAPass())
-            add!(pb, GlobalOptPass())
-            add!(pb, ResolveCPUReferencesPass())
-            add!(pb, GlobalDCEPass())
-            add!(pb, StripDeadPrototypesPass())
+        add!(pb, RecomputeGlobalsAAPass())
+        add!(pb, GlobalOptPass())
+        add!(pb, ResolveCPUReferencesPass())
+        add!(pb, GlobalDCEPass())
+        add!(pb, StripDeadPrototypesPass())
 
-            run!(pb, mod, llvm_machine(job.config.target))
-        end
-    else
-        @dispose pm=ModulePassManager() begin
-            global_optimizer!(pm)
-
-            resolve_cpu_references!(pm)
-
-            global_dce!(pm)
-            strip_dead_prototypes!(pm)
-
-            run!(pm, mod)
-        end
+        run!(pb, mod, llvm_machine(job.config.target))
     end
 
     return
@@ -78,12 +65,8 @@ function resolve_cpu_references!(mod::LLVM.Module)
 
     return changed
 end
-resolve_cpu_references!(pm::PassManager) =
-    add!(pm, ModulePass("ResolveCPUReferences", resolve_cpu_references!))
-if LLVM.has_newpm()
-    ResolveCPUReferencesPass() =
-        NewPMModulePass("ResolveCPUReferences", resolve_cpu_references!)
-end
+ResolveCPUReferencesPass() =
+    NewPMModulePass("ResolveCPUReferences", resolve_cpu_references!)
 
 
 function mcgen(@nospecialize(job::CompilerJob), mod::LLVM.Module, format=LLVM.API.LLVMAssemblyFile)

src/metal.jl

@@ -77,17 +77,10 @@ function finish_module!(@nospecialize(job::CompilerJob{MetalCompilerTarget}), mo
 
     # we emit properties (of the air and metal version) as private global constants,
     # so run the optimizer so that they are inlined before the rest of the optimizer runs.
-    if use_newpm
-        @dispose pb=NewPMPassBuilder() begin
-            add!(pb, RecomputeGlobalsAAPass())
-            add!(pb, GlobalOptPass())
-            run!(pb, mod)
-        end
-    else
-        @dispose pm=ModulePassManager() begin
-            global_optimizer!(pm)
-            run!(pm, mod)
-        end
+    @dispose pb=NewPMPassBuilder() begin
+        add!(pb, RecomputeGlobalsAAPass())
+        add!(pb, GlobalOptPass())
+        run!(pb, mod)
     end
 
     return functions(mod)[entry_fn]
@@ -121,22 +114,13 @@ function hide_noreturn!(mod::LLVM.Module)
     end
     any_noreturn || return false
 
-    if use_newpm
-        @dispose pb=NewPMPassBuilder() begin
-            add!(pb, AlwaysInlinerPass())
-            add!(pb, NewPMFunctionPassManager()) do fpm
-                add!(fpm, SimplifyCFGPass())
-                add!(fpm, InstCombinePass())
-            end
-            run!(pb, mod)
-        end
-    else
-        @dispose pm=ModulePassManager() begin
-            always_inliner!(pm)
-            cfgsimplification!(pm)
-            instruction_combining!(pm)
-            run!(pm, mod)
+    @dispose pb=NewPMPassBuilder() begin
+        add!(pb, AlwaysInlinerPass())
+        add!(pb, NewPMFunctionPassManager()) do fpm
+            add!(fpm, SimplifyCFGPass())
+            add!(fpm, InstCombinePass())
         end
+        run!(pb, mod)
     end
 
     return true
@@ -164,22 +148,13 @@ function finish_ir!(@nospecialize(job::CompilerJob{MetalCompilerTarget}), mod::L
     end
     if changed
         # lowering may have introduced additional functions marked `alwaysinline`
-        if use_newpm
-            @dispose pb=NewPMPassBuilder() begin
-                add!(pb, AlwaysInlinerPass())
-                add!(pb, NewPMFunctionPassManager()) do fpm
-                    add!(fpm, SimplifyCFGPass())
-                    add!(fpm, InstCombinePass())
-                end
-                run!(pb, mod)
-            end
-        else
-            @dispose pm=ModulePassManager() begin
-                always_inliner!(pm)
-                cfgsimplification!(pm)
-                instruction_combining!(pm)
-                run!(pm, mod)
+        @dispose pb=NewPMPassBuilder() begin
+            add!(pb, AlwaysInlinerPass())
+            add!(pb, NewPMFunctionPassManager()) do fpm
+                add!(fpm, SimplifyCFGPass())
+                add!(fpm, InstCombinePass())
             end
+            run!(pb, mod)
         end
     end
 
@@ -305,24 +280,13 @@ function add_address_spaces!(@nospecialize(job::CompilerJob), mod::LLVM.Module,
     LLVM.name!(new_f, fn)
 
     # clean-up after this pass (which runs after optimization)
-    if use_newpm
-       @dispose pb=NewPMPassBuilder() begin
-            add!(pb, SimplifyCFGPass())
-            add!(pb, SROAPass())
-            add!(pb, EarlyCSEPass())
-            add!(pb, InstCombinePass())
+    @dispose pb=NewPMPassBuilder() begin
+        add!(pb, SimplifyCFGPass())
+        add!(pb, SROAPass())
+        add!(pb, EarlyCSEPass())
+        add!(pb, InstCombinePass())
 
-            run!(pb, mod)
-        end
-    else
-        @dispose pm=ModulePassManager() begin
-            cfgsimplification!(pm)
-            scalar_repl_aggregates!(pm)
-            early_cse!(pm)
-            instruction_combining!(pm)
-
-            run!(pm, mod)
-        end
+        run!(pb, mod)
     end
 
     return new_f