Merge pull request #539 from JuliaGPU/tb/julia_ir_printer

Use the Julia LLVM IR printer to support the debuginfo argument.
JuliaGPU · Jan 31, 2020 · e4cbe2f · e4cbe2f
2 parents 9870462 + e66edf5
commit e4cbe2f
Show file tree

Hide file tree

Showing 3 changed files with 14 additions and 14 deletions.
diff --git a/src/reflection.jl b/src/reflection.jl
@@ -32,21 +32,21 @@ See also: [`@device_code_llvm`](@ref), InteractiveUtils.code_llvm
 """
 function code_llvm(io::IO, @nospecialize(func), @nospecialize(types);
                    cap::VersionNumber=current_capability(), kernel::Bool=false,
-                   optimize::Bool=true, raw::Bool=false,
+                   optimize::Bool=true, raw::Bool=false, debuginfo::Symbol=:default,
                    dump_module::Bool=false, strict::Bool=false, kwargs...)
     tt = Base.to_tuple_type(types)
     job = CompilerJob(func, tt, cap, kernel; kwargs...)
-    code_llvm(io, job; optimize=optimize,
-              raw=raw, dump_module=dump_module, strict=strict)
+    code_llvm(io, job; optimize=optimize, raw=raw, debuginfo=debuginfo,
+              dump_module=dump_module, strict=strict)
 end
 function code_llvm(io::IO, job::CompilerJob; optimize::Bool=true, raw::Bool=false,
-                   dump_module::Bool=false, strict::Bool=false)
-    ir, entry = codegen(:llvm, job; optimize=optimize, strip=!raw, strict=strict)
-    if dump_module
-        show(io, ir)
-    else
-        show(io, entry)
-    end
+                   debuginfo::Symbol=:default, dump_module::Bool=false, strict::Bool=false)
+    # NOTE: jl_dump_function_ir supports stripping metadata, so don't do it in the driver
+    ir, entry = codegen(:llvm, job; optimize=optimize, strip=false, strict=strict)
+    str = ccall(:jl_dump_function_ir, Ref{String},
+                (Ptr{Cvoid}, Bool, Bool, Ptr{UInt8}),
+                LLVM.ref(entry), !raw, dump_module, debuginfo)
+    print(io, str)
 end
 code_llvm(@nospecialize(func), @nospecialize(types); kwargs...) =
     code_llvm(stdout, func, types; kwargs...)

diff --git a/test/codegen.jl b/test/codegen.jl
@@ -152,13 +152,13 @@ end
     for f in (load, store)
         ir = sprint(io->CUDAnative.code_llvm(io, f,
                                              Tuple{CUDAnative.DevicePtr{Float32,AS.Global}};
-                                             dump_module=true))
+                                             dump_module=true, raw=true))
         @test occursin("ptxtbaa_global", ir)
 
         # no TBAA on generic pointers
         ir = sprint(io->CUDAnative.code_llvm(io, f,
                                              Tuple{CUDAnative.DevicePtr{Float32,AS.Generic}};
-                                             dump_module=true))
+                                             dump_module=true, raw=true))
         @test !occursin("ptxtbaa", ir)
     end
 
@@ -167,7 +167,7 @@ end
 
     ir = sprint(io->CUDAnative.code_llvm(io, cached_load,
                                          Tuple{CUDAnative.DevicePtr{Float32,AS.Global}};
-                                         dump_module=true))
+                                         dump_module=true, raw=true))
     @test occursin("ptxtbaa_global", ir)
 end
 

diff --git a/test/device/cuda.jl b/test/device/cuda.jl
@@ -20,7 +20,7 @@
 
     @testset "range metadata" begin
         foobar() = threadIdx().x
-        ir = sprint(io->CUDAnative.code_llvm(io, foobar, Tuple{}))
+        ir = sprint(io->CUDAnative.code_llvm(io, foobar, Tuple{}; raw=true))
 
         @test occursin(r"call .+ @llvm.nvvm.read.ptx.sreg.tid.x.+ !range", ir)
     end