Use NewPM wherever possible (#482)

Co-authored-by: Tim Besard <tim.besard@gmail.com>

Manifest.toml

@@ -2,7 +2,7 @@
 
 julia_version = "1.8.5"
 manifest_format = "2.0"
-project_hash = "27e7a553e90a68d1f177019877f2e4b06a23c81a"
+project_hash = "7fba634d9208dc361f8caa397a5e8d2621e56bcf"
 
 [[deps.ArgTools]]
 uuid = "0dad84c5-d112-42e6-8d28-ef12dabb789f"
@@ -48,15 +48,15 @@ version = "1.5.0"
 
 [[deps.LLVM]]
 deps = ["CEnum", "LLVMExtra_jll", "Libdl", "Printf", "Unicode"]
-git-tree-sha1 = "8695a49bfe05a2dc0feeefd06b4ca6361a018729"
+git-tree-sha1 = "f7e39b1ecd9531475bbf3b25363027ba14c3e563"
 uuid = "929cbde3-209d-540e-8aea-75f648917ca0"
-version = "6.1.0"
+version = "6.2.0"
 
 [[deps.LLVMExtra_jll]]
 deps = ["Artifacts", "JLLWrappers", "LazyArtifacts", "Libdl", "TOML"]
-git-tree-sha1 = "c35203c1e1002747da220ffc3c0762ce7754b08c"
+git-tree-sha1 = "7ca6850ae880cc99b59b88517545f91a52020afa"
 uuid = "dad2f222-ce93-54a1-a47d-0025e8a3acab"
-version = "0.0.23+0"
+version = "0.0.25+0"
 
 [[deps.LazyArtifacts]]
 deps = ["Artifacts", "Pkg"]

Project.toml

@@ -15,7 +15,7 @@ UUIDs = "cf7118a7-6976-5b1a-9a39-7adc72f591a4"
 
 [compat]
 ExprTools = "0.1"
-LLVM = "6"
+LLVM = "6.2"
 Scratch = "1"
 TimerOutputs = "0.5"
 julia = "1.8"

src/GPUCompiler.jl

@@ -14,6 +14,8 @@ using Scratch: @get_scratch!
 const CC = Core.Compiler
 using Core: MethodInstance, CodeInstance, CodeInfo
 
+const use_newpm = LLVM.has_newpm()
+
 include("utils.jl")
 
 # compiler interface and implementations

src/driver.jl

@@ -313,6 +313,7 @@ const __llvm_initialized = Ref(false)
         # global variables. this makes sure that the optimizer can, e.g.,
         # rewrite function signatures.
         if toplevel
+            # TODO: there's no good API to use internalize with the new pass manager yet
             @dispose pm=ModulePassManager() begin
                 exports = collect(values(jobs))
                 for gvar in globals(ir)
@@ -340,20 +341,37 @@ const __llvm_initialized = Ref(false)
                 # deferred codegen has some special optimization requirements,
                 # which also need to happen _after_ regular optimization.
                 # XXX: make these part of the optimizer pipeline?
-                has_deferred_jobs && @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)
+                if has_deferred_jobs
+                    if use_newpm
+                        @dispose pb=PassBuilder() mpm=NewPMModulePassManager(pb) begin
+                            add!(mpm, NewPMFunctionPassManager) do fpm
+                                add!(fpm, InstCombinePass())
+                            end
+                            add!(mpm, AlwaysInlinerPass())
+                            add!(mpm, NewPMFunctionPassManager) do fpm
+                                add!(fpm, SROAPass())
+                                add!(fpm, GVNPass())
+                            end
+                            add!(mpm, MergeFunctionsPass())
+                            run!(mpm, ir)
+                        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)
+                        end
+                    end
                 end
             end
 
@@ -363,18 +381,29 @@ const __llvm_initialized = Ref(false)
 
         if cleanup
             @timeit_debug to "clean-up" begin
-                # 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)
+                if use_newpm
+                    @dispose pb=PassBuilder() mpm=NewPMModulePassManager(pb) begin
+                        add!(mpm, RecomputeGlobalsAAPass())
+                        add!(mpm, GlobalOptPass())
+                        add!(mpm, GlobalDCEPass())
+                        add!(mpm, StripDeadPrototypesPass())
+                        add!(mpm, ConstantMergePass())
+                        run!(mpm, ir)
+                    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
                 end
             end
         end

src/gcn.jl

@@ -37,10 +37,7 @@ isintrinsic(::CompilerJob{GCNCompilerTarget}, fn::String) = in(fn, gcn_intrinsic
 
 function finish_module!(@nospecialize(job::CompilerJob{GCNCompilerTarget}),
                         mod::LLVM.Module, entry::LLVM.Function)
-    @dispose pm=ModulePassManager() begin
-        add!(pm, ModulePass("LowerThrowExtra", lower_throw_extra!))
-        run!(pm, mod)
-    end
+    lower_throw_extra!(mod)
 
     if job.config.kernel
         # calling convention

src/irgen.jl

@@ -96,6 +96,7 @@ function irgen(@nospecialize(job::CompilerJob))
             end
         end
 
+        # TODO: there's no good API to use internalize with the new pass manager yet
         @dispose pm=ModulePassManager() begin
             global current_job
             current_job = job

src/mcgen.jl

@@ -3,18 +3,30 @@
 # final preparations for the module to be compiled to machine code
 # these passes should not be run when e.g. compiling to write to disk.
 function prepare_execution!(@nospecialize(job::CompilerJob), mod::LLVM.Module)
-    @dispose pm=ModulePassManager() begin
-        global current_job
-        current_job = job
+    global current_job
+    current_job = job
 
-        global_optimizer!(pm)
+    if use_newpm
+        @dispose pb=PassBuilder() mpm=NewPMModulePassManager(pb) begin
+            add!(mpm, RecomputeGlobalsAAPass())
+            add!(mpm, GlobalOptPass())
+            resolve_cpu_references!(mod)
+            add!(legacy2newpm(resolve_cpu_references!), mpm)
+            add!(mpm, GlobalDCEPass())
+            add!(mpm, StripDeadPrototypesPass())
+            run!(mpm, mod)
+        end
+    else
+        @dispose pm=ModulePassManager() begin
+            global_optimizer!(pm)
 
-        add!(pm, ModulePass("ResolveCPUReferences", resolve_cpu_references!))
+            add!(pm, ModulePass("ResolveCPUReferences", resolve_cpu_references!))
 
-        global_dce!(pm)
-        strip_dead_prototypes!(pm)
+            global_dce!(pm)
+            strip_dead_prototypes!(pm)
 
-        run!(pm, mod)
+            run!(pm, mod)
+        end
     end
 
     return

src/metal.jl

@@ -87,10 +87,19 @@ function finish_module!(@nospecialize(job::CompilerJob{MetalCompilerTarget}), mo
     # add metadata to AIR intrinsics LLVM doesn't know about
     annotate_air_intrinsics!(job, mod)
 
-    @dispose pm=ModulePassManager() begin
-        # we emit properties (of the device and ptx isa) as private global constants,
-        # so run the optimizer so that they are inlined before the rest of the optimizer runs.
-        global_optimizer!(pm)
+    # 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=PassBuilder() mpm=NewPMModulePassManager(pb) begin
+            add!(mpm, RecomputeGlobalsAAPass())
+            add!(mpm, GlobalOptPass())
+            run!(mpm, mod)
+        end
+    else
+        @dispose pm=ModulePassManager() begin
+            global_optimizer!(pm)
+            run!(pm, mod)
+        end
     end
 
     return functions(mod)[entry_fn]
@@ -121,11 +130,22 @@ function finish_ir!(@nospecialize(job::CompilerJob{MetalCompilerTarget}), mod::L
     end
     if changed
         # lowering may have introduced additional functions marked `alwaysinline`
-        @dispose pm=ModulePassManager() begin
-            always_inliner!(pm)
-            cfgsimplification!(pm)
-            instruction_combining!(pm)
-            run!(pm, mod)
+        if use_newpm
+            @dispose pb=PassBuilder() mpm=NewPMModulePassManager(pb) begin
+                add!(mpm, AlwaysInlinerPass())
+                add!(mpm, NewPMFunctionPassManager) do fpm
+                    add!(fpm, SimplifyCFGPass())
+                    add!(fpm, InstCombinePass())
+                end
+                run!(mpm, mod)
+            end
+        else
+            @dispose pm=ModulePassManager() begin
+                always_inliner!(pm)
+                cfgsimplification!(pm)
+                instruction_combining!(pm)
+                run!(pm, mod)
+            end
         end
     end
 
@@ -158,11 +178,22 @@ end
         end
 
         if any_noreturn
-            @dispose pm=ModulePassManager() begin
-                always_inliner!(pm)
-                cfgsimplification!(pm)
-                instruction_combining!(pm)
-                run!(pm, mod)
+            if use_newpm
+                @dispose pb=PassBuilder() mpm=NewPMModulePassManager(pb) begin
+                    add!(mpm, AlwaysInlinerPass())
+                    add!(mpm, NewPMFunctionPassManager) do fpm
+                        add!(fpm, SimplifyCFGPass())
+                        add!(fpm, InstCombinePass())
+                    end
+                    run!(mpm, mod)
+                end
+            else
+                @dispose pm=ModulePassManager() begin
+                    always_inliner!(pm)
+                    cfgsimplification!(pm)
+                    instruction_combining!(pm)
+                    run!(pm, mod)
+                end
             end
         end
     end
@@ -294,13 +325,26 @@ function add_address_spaces!(@nospecialize(job::CompilerJob), mod::LLVM.Module,
     LLVM.name!(new_f, fn)
 
     # clean-up after this pass (which runs after optimization)
-    @dispose pm=ModulePassManager() begin
-        cfgsimplification!(pm)
-        scalar_repl_aggregates!(pm)
-        early_cse!(pm)
-        instruction_combining!(pm)
+    if use_newpm
+       @dispose pb=PassBuilder() mpm=NewPMModulePassManager(pb) begin
+            add!(mpm, NewPMFunctionPassManager) do fpm
+                add!(fpm, SimplifyCFGPass())
+                add!(fpm, SROAPass())
+                add!(fpm, EarlyCSEPass())
+                add!(fpm, InstCombinePass())
+            end
 
-        run!(pm, mod)
+            run!(mpm, mod)
+        end
+    else
+        @dispose pm=ModulePassManager() begin
+            cfgsimplification!(pm)
+            scalar_repl_aggregates!(pm)
+            early_cse!(pm)
+            instruction_combining!(pm)
+
+            run!(pm, mod)
+        end
     end
 
     return new_f