Update New PassManager Pipeline (#559)

src/optim.jl

@@ -38,9 +38,10 @@ function buildNewPMPipeline!(mpm, @nospecialize(job::CompilerJob), opt_level=2)
  add!(mpm, NewPMFunctionPassManager) do fpm
  buildLoopOptimizerPipeline(fpm, job, opt_level)
  buildScalarOptimizerPipeline(fpm, job, opt_level)
- if false && opt_level >= 2
+ if uses_julia_runtime(job) && opt_level >= 2
  # XXX: we disable vectorization, as this generally isn't useful for GPU targets
  # and actually causes issues with some back-end compilers (like Metal).
+ # TODO: Make this not dependent on `uses_julia_runtime` (likely CPU), but it's own control
  buildVectorPipeline(fpm, job, opt_level)
  end
  if isdebug(:optim)
@@ -112,6 +113,8 @@ end
 
 function buildLoopOptimizerPipeline(fpm, @nospecialize(job::CompilerJob), opt_level)
  add!(fpm, NewPMLoopPassManager) do lpm
+ # TODO LowerSIMDLoopPass
+ # LoopPass since JuliaLang/julia#51883
  if opt_level >= 2
  add!(lpm, LoopRotatePass())
  end
@@ -121,7 +124,7 @@ function buildLoopOptimizerPipeline(fpm, @nospecialize(job::CompilerJob), opt_le
  add!(fpm, NewPMLoopPassManager, #=UseMemorySSA=#true) do lpm
  add!(lpm, LICMPass())
  add!(lpm, JuliaLICMPass())
- add!(lpm, SimpleLoopUnswitchPass())
+ add!(lpm, SimpleLoopUnswitchPass(SimpleLoopUnswitchPassOptions(nontrivial=true, trivial=true)))
  add!(lpm, LICMPass())
  add!(lpm, JuliaLICMPass())
  end
@@ -186,23 +189,41 @@ function buildVectorPipeline(fpm, @nospecialize(job::CompilerJob), opt_level)
 end
 
 function buildIntrinsicLoweringPipeline(mpm, @nospecialize(job::CompilerJob), opt_level)
- # lower exception handling
- if uses_julia_runtime(job)
+ add!(mpm, RemoveNIPass())
+
+ # lower GC intrinsics
+ add!(mpm, NewPMFunctionPassManager) do fpm
+ if !uses_julia_runtime(job)
+ add!(legacy2newpm(lower_gc_frame!), fpm)
+ end
+ end
+
+ # lower kernel state intrinsics
+ # NOTE: we can only do so here, as GC lowering can introduce calls to the runtime,
+ # and thus additional uses of the kernel state intrinsics.
+ if job.config.kernel
+ # TODO: now that all kernel state-related passes are being run here, merge some?
+ add!(legacy2newpm(add_kernel_state!), mpm)
  add!(mpm, NewPMFunctionPassManager) do fpm
-  add!(fpm, LowerExcHandlersPass())
+ add!(legacy2newpm(lower_kernel_state!), fpm)
  end
+ add!(legacy2newpm(cleanup_kernel_state!), mpm)
  end
 
- add!(mpm, NewPMFunctionPassManager) do fpm
- add!(fpm, GCInvariantVerifierPass())
+ if !uses_julia_runtime(job)
+ # remove dead uses of ptls
+ add!(mpm, NewPMFunctionPassManager) do fpm
+ add!(fpm, ADCEPass())
+ end
+ add!(legacy2newpm(lower_ptls!), mpm)
  end
- add!(mpm, RemoveNIPass())
 
- # lower GC intrinsics
  add!(mpm, NewPMFunctionPassManager) do fpm
- if !uses_julia_runtime(job)
- add!(legacy2newpm(lower_gc_frame!), fpm)
+ # lower exception handling
+ if uses_julia_runtime(job)
+ add!(fpm, LowerExcHandlersPass())
  end
+ add!(fpm, GCInvariantVerifierPass())
  add!(fpm, LateLowerGCPass())
  if uses_julia_runtime(job) && VERSION >= v"1.11.0-DEV.208"
  add!(fpm, FinalLowerGCPass())
@@ -220,27 +241,9 @@ function buildIntrinsicLoweringPipeline(mpm, @nospecialize(job::CompilerJob), op
  end
  end
 
- # lower kernel state intrinsics
- # NOTE: we can only do so here, as GC lowering can introduce calls to the runtime,
- # and thus additional uses of the kernel state intrinsics.
- if job.config.kernel
- # TODO: now that all kernel state-related passes are being run here, merge some?
- add!(legacy2newpm(add_kernel_state!), mpm)
- add!(mpm, NewPMFunctionPassManager) do fpm
- add!(legacy2newpm(lower_kernel_state!), fpm)
- end
- add!(legacy2newpm(cleanup_kernel_state!), mpm)
- end
-
  # lower PTLS intrinsics
  if uses_julia_runtime(job)
  add!(mpm, LowerPTLSPass())
- else
- # remove dead uses of ptls
- add!(mpm, NewPMFunctionPassManager) do fpm
- add!(fpm, ADCEPass())
- end
- add!(legacy2newpm(lower_ptls!), mpm)
  end
 
  if opt_level >= 1

test/gcn_tests.jl

@@ -191,6 +191,10 @@ false && @testset "GC and TLS lowering" begin
 
  asm = sprint(io->GCN.code_native(io, mod.kernel, Tuple{Int}))
  @test occursin("gpu_gc_pool_alloc", asm)
+ @test !occursin("julia.push_gc_frame", asm)
+ @test !occursin("julia.pop_gc_frame", asm)
+ @test !occursin("julia.get_gc_frame_slot", asm)
+ @test !occursin("julia.new_gc_frame", asm)
 
  # make sure that we can still ellide allocations
  function ref_kernel(ptr, i)

test/ptx_tests.jl

@@ -279,6 +279,11 @@ end
 
  asm = sprint(io->PTX.code_native(io, mod.kernel, Tuple{Int}))
  @test occursin("gpu_gc_pool_alloc", asm)
+ @test !occursin("julia.push_gc_frame", asm)
+ @test !occursin("julia.pop_gc_frame", asm)
+ @test !occursin("julia.get_gc_frame_slot", asm)
+ @test !occursin("julia.new_gc_frame", asm)
+
 
  # make sure that we can still ellide allocations
  function ref_kernel(ptr, i)