Adapt to LLVM 17. (#583)

Author: maleadt

Project.toml

@@ -19,7 +19,7 @@ InteractiveUtils = "1"
 Libdl = "1"
 Logging = "1"
 UUIDs = "1"
-LLVM = "6.6"
+LLVM = "7.1"
 Scratch = "1"
 TimerOutputs = "0.5"
 julia = "1.8"

src/driver.jl

@@ -325,16 +325,22 @@ 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)
-                    if linkage(gvar) == LLVM.API.LLVMExternalLinkage
-                        push!(exports, LLVM.name(gvar))
-                    end
+            preserved_gvs = collect(values(jobs))
+            for gvar in globals(ir)
+                if linkage(gvar) == LLVM.API.LLVMExternalLinkage
+                    push!(preserved_gvs, LLVM.name(gvar))
+                end
+            end
+            if use_newpm && LLVM.version() >= v"17"
+                @dispose pb=PassBuilder() mpm=NewPMModulePassManager(pb) begin
+                    add!(mpm, InternalizePass(InternalizePassOptions(; preserved_gvs)))
+                    run!(mpm, ir)
+                end
+            else
+                @dispose pm=ModulePassManager() begin
+                    internalize!(pm, preserved_gvs)
+                    run!(pm, ir)
                 end
-                internalize!(pm, exports)
-                run!(pm, ir)
             end
         end
 

src/irgen.jl

@@ -95,27 +95,29 @@ 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
-
-            linkage!(entry, LLVM.API.LLVMExternalLinkage)
-
-            # internalize all functions, but keep exported global variables
-            exports = String[LLVM.name(entry)]
-            for gvar in globals(mod)
-                push!(exports, LLVM.name(gvar))
+        # internalize all functions and, but keep exported global variables.
+        linkage!(entry, LLVM.API.LLVMExternalLinkage)
+        preserved_gvs = String[LLVM.name(entry)]
+        for gvar in globals(mod)
+            push!(preserved_gvs, LLVM.name(gvar))
+        end
+        if use_newpm && LLVM.version() >= v"17"
+            @dispose pb=PassBuilder() mpm=NewPMModulePassManager(pb) begin
+                add!(mpm, InternalizePass(InternalizePassOptions(; preserved_gvs)))
+                add!(mpm, AlwaysInlinerPass())
+                run!(mpm, mod)
+            end
+        else
+            @dispose pm=ModulePassManager() begin
+                internalize!(pm, preserved_gvs)
+                always_inliner!(pm)
+                run!(pm, mod)
             end
-            internalize!(pm, exports)
-
-            # inline llvmcall bodies
-            always_inliner!(pm)
-
-            can_throw(job) || add!(pm, ModulePass("LowerThrow", lower_throw!))
-
-            run!(pm, mod)
         end
+
+        global current_job
+        current_job = job
+        can_throw(job) || lower_throw!(mod)
     end
 
     return mod, compiled

src/optim.jl

@@ -13,7 +13,6 @@ end
 ## new pm
 
 function optimize_newpm!(@nospecialize(job::CompilerJob), mod::LLVM.Module; opt_level)
-    triple = llvm_triple(job.config.target)
     tm = llvm_machine(job.config.target)
 
     global current_job
@@ -34,7 +33,9 @@ function buildNewPMPipeline!(mpm, @nospecialize(job::CompilerJob), opt_level)
     buildEarlySimplificationPipeline(mpm, job, opt_level)
     add!(mpm, AlwaysInlinerPass())
     buildEarlyOptimizerPipeline(mpm, job, opt_level)
-    add!(mpm, LowerSIMDLoopPass())
+    if VERSION < v"1.10"
+        add!(mpm, LowerSIMDLoopPass())
+    end
     add!(mpm, NewPMFunctionPassManager) do fpm
         buildLoopOptimizerPipeline(fpm, job, opt_level)
         buildScalarOptimizerPipeline(fpm, job, opt_level)
@@ -113,8 +114,9 @@ end
 
 function buildLoopOptimizerPipeline(fpm, @nospecialize(job::CompilerJob), opt_level)
     add!(fpm, NewPMLoopPassManager) do lpm
-        # TODO LowerSIMDLoopPass
-        # LoopPass since JuliaLang/julia#51883
+        if VERSION >= v"1.10"
+            add!(lpm, LowerSIMDLoopPass())
+        end
         if opt_level >= 2
             add!(lpm, LoopRotatePass())
         end

src/ptx.jl

@@ -153,42 +153,86 @@ end
 function optimize_module!(@nospecialize(job::CompilerJob{PTXCompilerTarget}),
                           mod::LLVM.Module)
     tm = llvm_machine(job.config.target)
-    # TODO can't convert to newpm because speculative-execution doesn't have a parameter in the default PassBuilder parser
-    @dispose pm=ModulePassManager() begin
-        add_library_info!(pm, triple(mod))
-        add_transform_info!(pm, tm)
-
-        # TODO: need to run this earlier; optimize_module! is called after addOptimizationPasses!
-        add!(pm, FunctionPass("NVVMReflect", nvvm_reflect!))
-
-        # needed by GemmKernels.jl-like code
-        speculative_execution_if_has_branch_divergence!(pm)
+    # TODO: Use the registered target passes (JuliaGPU/GPUCompiler.jl#450)
+    if use_newpm
+        @dispose pb=PassBuilder(tm) mpm=NewPMModulePassManager(pb) begin
+            add!(mpm, NewPMFunctionPassManager) do fpm
+                # TODO: need to run this earlier; optimize_module! is called after addOptimizationPasses!
+                add!(legacy2newpm(nvvm_reflect!), fpm)
+
+                # needed by GemmKernels.jl-like code
+                add!(fpm, SpeculativeExecutionPass())
+
+                # NVPTX's target machine info enables runtime unrolling,
+                # but Julia's pass sequence only invokes the simple unroller.
+                add!(fpm, LoopUnrollPass(LoopUnrollOptions(; job.config.opt_level)))
+                add!(fpm, InstCombinePass())        # clean-up redundancy
+                add!(fpm, NewPMLoopPassManager) do lpm
+                    add!(lpm, LICMPass())           # the inner runtime check might be
+                                                    # outer loop invariant
+                end
+
+                # the above loop unroll pass might have unrolled regular, non-runtime nested loops.
+                # that code still needs to be optimized (arguably, multiple unroll passes should be
+                # scheduled by the Julia optimizer). do so here, instead of re-optimizing entirely.
+                if job.config.opt_level == 2
+                    add!(fpm, GVNPass())
+                elseif job.config.opt_level == 1
+                    add!(fpm, EarlyCSEPass())
+                end
+                add!(fpm, DSEPass())
+
+                add!(fpm, SimplifyCFGPass())
+            end
 
-        # NVPTX's target machine info enables runtime unrolling,
-        # but Julia's pass sequence only invokes the simple unroller.
-        loop_unroll!(pm)
-        instruction_combining!(pm)  # clean-up redundancy
-        licm!(pm)                   # the inner runtime check might be outer loop invariant
+            # get rid of the internalized functions; now possible unused
+            add!(mpm, GlobalDCEPass())
 
-        # the above loop unroll pass might have unrolled regular, non-runtime nested loops.
-        # that code still needs to be optimized (arguably, multiple unroll passes should be
-        # scheduled by the Julia optimizer). do so here, instead of re-optimizing entirely.
-        early_csemem_ssa!(pm) # TODO: gvn instead? see NVPTXTargetMachine.cpp::addEarlyCSEOrGVNPass
-        dead_store_elimination!(pm)
+            run!(mpm, mod, tm)
+        end
+    else
+        @dispose pm=ModulePassManager() begin
+            add_library_info!(pm, triple(mod))
+            add_transform_info!(pm, tm)
+
+            # TODO: need to run this earlier; optimize_module! is called after addOptimizationPasses!
+            add!(pm, FunctionPass("NVVMReflect", nvvm_reflect!))
+
+            # needed by GemmKernels.jl-like code
+            speculative_execution_if_has_branch_divergence!(pm)
+
+            # NVPTX's target machine info enables runtime unrolling,
+            # but Julia's pass sequence only invokes the simple unroller.
+            loop_unroll!(pm)
+            instruction_combining!(pm)  # clean-up redundancy
+            licm!(pm)                   # the inner runtime check might be outer loop invariant
+
+            # the above loop unroll pass might have unrolled regular, non-runtime nested loops.
+            # that code still needs to be optimized (arguably, multiple unroll passes should be
+            # scheduled by the Julia optimizer). do so here, instead of re-optimizing entirely.
+            if job.config.opt_level == 2
+                gvn!(pm)
+            elseif job.config.opt_level == 1
+                early_cse!(pm)
+            end
+            dead_store_elimination!(pm)
 
-        cfgsimplification!(pm)
+            cfgsimplification!(pm)
 
-        # get rid of the internalized functions; now possible unused
-        global_dce!(pm)
+            # get rid of the internalized functions; now possible unused
+            global_dce!(pm)
 
-        run!(pm, mod)
+            run!(pm, mod)
+        end
     end
 end
 
 function finish_ir!(@nospecialize(job::CompilerJob{PTXCompilerTarget}),
                     mod::LLVM.Module, entry::LLVM.Function)
-    for f in functions(mod)
-        lower_unreachable!(f)
+    if LLVM.version() < v"17"
+        for f in functions(mod)
+            lower_unreachable!(f)
+        end
     end
 
     if job.config.kernel

test/ptx_tests.jl

@@ -114,7 +114,7 @@ end
     @test occursin(r"@\w*kernel\w*\(\[1 x i64\] %state", ir)
 
     # child1 doesn't use the state
-    @test occursin(r"@\w*child1\w*\((i64|i8\*)", ir)
+    @test occursin(r"@\w*child1\w*\((i64|i8\*|ptr)", ir)
 
     # child2 does
     @test occursin(r"@\w*child2\w*\(\[1 x i64\] %state", ir)
@@ -341,7 +341,7 @@ precompile_test_harness("Inference caching") do load_path
             job, _ = PTXCompiler.create_job(kernel, ())
             GPUCompiler.code_typed(job)
         end
-        
+
         # identity is foreign
         @setup_workload begin
             job, _ = PTXCompiler.create_job(identity, (Int,))