Add imaging_mode support (static compilation)

Author: jpsamaroo

src/driver.jl

@@ -38,7 +38,8 @@ function compile(target::Symbol, @nospecialize(job::CompilerJob);
 
     return codegen(target, job;
                    libraries=libraries, deferred_codegen=deferred_codegen,
-                   optimize=optimize, strip=strip, validate=validate, only_entry=only_entry)
+                   optimize=optimize, strip=strip, validate=validate,
+                   only_entry=only_entry)
 end
 
 # primitive mechanism for deferred compilation, for implementing CUDA dynamic parallelism.

src/gcn.jl

@@ -10,7 +10,7 @@ end
 
 llvm_triple(::GCNCompilerTarget) = "amdgcn-amd-amdhsa"
 
-function llvm_machine(target::GCNCompilerTarget)
+function llvm_machine(target::GCNCompilerTarget, static)
     triple = llvm_triple(target)
     t = Target(triple=triple)
 
@@ -65,7 +65,7 @@ function optimize_module!(job::CompilerJob{GCNCompilerTarget}, mod::LLVM.Module)
     triple!(mod, llvm_triple(job.target))
     datalayout!(mod, llvm_datalayout(job.target))
 
-    tm = llvm_machine(job.target)
+    tm = llvm_machine(job.target, job.source.static)
     ModulePassManager() do pm
         add_library_info!(pm, triple(mod))
         add_transform_info!(pm, tm)

src/interface.jl

@@ -23,18 +23,19 @@ source_code(::AbstractCompilerTarget) = "text"
 
 llvm_triple(::AbstractCompilerTarget) = error("Not implemented")
 
-function llvm_machine(target::AbstractCompilerTarget)
+function llvm_machine(target::AbstractCompilerTarget, static)
     triple = llvm_triple(target)
 
     t = Target(triple=triple)
 
-    tm = TargetMachine(t, triple)
+    reloc = static ? LLVM.API.LLVMRelocPIC : LLVM.API.LLVMRelocDefault
+    tm = TargetMachine(t, triple, "", "", LLVM.API.LLVMCodeGenLevelDefault, reloc)
     asm_verbosity!(tm, true)
 
     return tm
 end
 
-llvm_datalayout(target::AbstractCompilerTarget) = DataLayout(llvm_machine(target))
+llvm_datalayout(target::AbstractCompilerTarget) = DataLayout(llvm_machine(target, false))
 
 
 ## params
@@ -57,12 +58,13 @@ struct FunctionSpec{F,TT}
     tt::DataType
     kernel::Bool
     name::Union{Nothing,String}
+    static::Bool
 end
 
 # put the function and argument types in typevars
 # so that we can access it from generated functions
-FunctionSpec(f, tt=Tuple{}, kernel=true, name=nothing) =
-    FunctionSpec{typeof(f),tt}(f, tt, kernel, name)
+FunctionSpec(f, tt=Tuple{}, kernel=true, name=nothing, static=false) =
+    FunctionSpec{typeof(f),tt}(f, tt, kernel, name, static)
 
 function signature(spec::FunctionSpec)
     fn = something(spec.name, nameof(spec.f))

src/jlgen.jl

@@ -214,9 +214,10 @@ function compile_method_instance(@nospecialize(job::CompilerJob), method_instanc
     end
 
     # generate IR
+    extern_policy = job.source.static ? 2 : 1
     native_code = ccall(:jl_create_native, Ptr{Cvoid},
                         (Vector{MethodInstance}, Base.CodegenParams, Cint),
-                        [method_instance], params, #=extern policy=# 1)
+                        [method_instance], params, extern_policy)
     @assert native_code != C_NULL
     llvm_mod_ref = ccall(:jl_get_llvm_module, LLVM.API.LLVMModuleRef,
                          (Ptr{Cvoid},), native_code)

src/mcgen.jl

@@ -69,7 +69,7 @@ end
 
 
 function mcgen(@nospecialize(job::CompilerJob), mod::LLVM.Module, f::LLVM.Function, format=LLVM.API.LLVMAssemblyFile)
-    tm = llvm_machine(job.target)
+    tm = llvm_machine(job.target, job.source.static)
 
     return String(emit(tm, mod, format))
 end

src/native.jl

@@ -11,12 +11,14 @@ end
 
 llvm_triple(::NativeCompilerTarget) = Sys.MACHINE
 
-function llvm_machine(target::NativeCompilerTarget)
+function llvm_machine(target::NativeCompilerTarget, static)
     triple = llvm_triple(target)
 
     t = Target(triple=triple)
 
-    tm = TargetMachine(t, triple, target.cpu, target.features)
+    optlevel = LLVM.API.LLVMCodeGenLevelDefault
+    reloc = static ? LLVM.API.LLVMRelocPIC : LLVM.API.LLVMRelocDefault
+    tm = TargetMachine(t, triple, target.cpu, target.features, optlevel, reloc)
     asm_verbosity!(tm, true)
 
     return tm

src/optim.jl

@@ -1,7 +1,7 @@
 # LLVM IR optimization
 
 function optimize!(@nospecialize(job::CompilerJob), mod::LLVM.Module)
-    tm = llvm_machine(job.target)
+    tm = llvm_machine(job.target, job.source.static)
 
     function initialize!(pm)
         add_library_info!(pm, triple(mod))

src/ptx.jl

@@ -18,13 +18,15 @@ source_code(::PTXCompilerTarget) = "ptx"
 
 llvm_triple(::PTXCompilerTarget) = Int===Int64 ? "nvptx64-nvidia-cuda" : "nvptx-nvidia-cuda"
 
-function llvm_machine(target::PTXCompilerTarget)
+function llvm_machine(target::PTXCompilerTarget, static)
     triple = llvm_triple(target)
     t = Target(triple=triple)
 
     cpu = "sm_$(target.cap.major)$(target.cap.minor)"
     feat = "+ptx60" # we only support CUDA 9.0+ and LLVM 6.0+
-    tm = TargetMachine(t, triple, cpu, feat)
+    optlevel = LLVM.API.LLVMCodeGenLevelDefault
+    reloc = static ? LLVM.API.LLVMRelocPIC : LLVM.API.LLVMRelocDefault
+    tm = TargetMachine(t, triple, cpu, feat, optlevel, reloc)
     asm_verbosity!(tm, true)
 
     return tm
@@ -127,7 +129,7 @@ function add_lowering_passes!(job::CompilerJob{PTXCompilerTarget}, pm::LLVM.Pass
 end
 
 function optimize_module!(job::CompilerJob{PTXCompilerTarget}, mod::LLVM.Module)
-    tm = llvm_machine(job.target)
+    tm = llvm_machine(job.target, job.source.static)
     ModulePassManager() do pm
         add_library_info!(pm, triple(mod))
         add_transform_info!(pm, tm)

src/spirv.jl

@@ -18,7 +18,7 @@ end
 llvm_triple(::SPIRVCompilerTarget) = Int===Int64 ? "spir64-unknown-unknown" : "spirv-unknown-unknown"
 
 # SPIRV is not supported by our LLVM builds, so we can't get a target machine
-llvm_machine(::SPIRVCompilerTarget) = nothing
+llvm_machine(::SPIRVCompilerTarget, static) = nothing
 
 llvm_datalayout(::SPIRVCompilerTarget) = Int===Int64 ?
     "e-i64:64-v16:16-v24:32-v32:32-v48:64-v96:128-v192:256-v256:256-v512:512-v1024:1024" :