switch to using clang instead of bultin LLVM for compilation

Author: simeonschaub

src/compiler/codegen.jl

@@ -42,6 +42,44 @@ function GPUCompiler.link_libraries!(@nospecialize(job::HIPCompilerJob), mod::LL
         wavefrontsize64=job.config.params.wavefrontsize64)
 end
 
+
+GPUCompiler.@unlocked function GPUCompiler.mcgen(
+    @nospecialize(job::CompilerJob{<:GCNCompilerTarget}), mod::LLVM.Module, format=LLVM.API.LLVMAssemblyFile,
+)
+    clang_path = AMDGPU.ROCmDiscovery.clang_path
+    if isempty(clang_path)
+        # Fallback to GPUCompiler default if no external clang is found
+        return invoke(GPUCompiler.mcgen, Tuple{CompilerJob, LLVM.Module, typeof(format)}, job, mod, format)
+    end
+
+    target = job.config.target
+    filetype = if format == LLVM.API.LLVMAssemblyFile
+        "asm"
+    elseif format == LLVM.API.LLVMObjectFile
+        "obj"
+    else
+        error("Unsupported GCN output format $format")
+    end
+
+    input  = tempname(cleanup=false) * ".bc"
+    output = tempname(cleanup=false) * (filetype == "asm" ? ".s" : ".o")
+    write(input, mod)
+
+    cmd = `$clang_path -x ir $input -mcpu=$(target.dev_isa) --target=$(GPUCompiler.llvm_triple(target)) -nogpulib $(filetype == "asm" ? "-S" : "-c") -o $output`
+
+    try
+        run(cmd)
+    catch
+        error("""Failed to compile to GCN with external clang.
+                 If you think this is a bug, please file an issue and attach $(input).""")
+    end
+
+    code = filetype == "asm" ? read(output, String) : String(read(output))
+    rm(input)
+    rm(output)
+    return code
+end
+
 function GPUCompiler.finish_module!(
     @nospecialize(job::HIPCompilerJob), mod::LLVM.Module, entry::LLVM.Function,
 )

src/compiler/device_libs.jl

@@ -28,6 +28,9 @@ function link_device_libs!(
     wavefrontsize64::Bool,
 )
     isnothing(libdevice_libs) && return
+    if !isempty(AMDGPU.ROCmDiscovery.clang_path)
+        return
+    end
 
     # 1. Load other libraries.
     lib_names = ("hc", "hip", "irif", "ockl", "opencl", "ocml")

src/discovery/discovery.jl

@@ -1,6 +1,6 @@
 module ROCmDiscovery
 
-export lld_artifact, lld_path, libhsaruntime, libdevice_libs, libhip
+export lld_artifact, lld_path, clang_path, clang_artifact, libhsaruntime, libdevice_libs, libhip
 export librocblas, librocsparse, librocsolver
 export librocrand, librocfft, libMIOpen_path
 
@@ -26,6 +26,16 @@ function get_ld_lld(rocm_path::String)::Tuple{String, Bool}
     return (LLD_jll.lld_path, true)
 end
 
+function get_clang(rocm_path::String)::Tuple{String, Bool}
+    clang_path = find_clang(rocm_path)
+    if isempty(clang_path)
+        # Not found, could use LLD_jll or something if it had clang, but it doesn't.
+        # So return empty
+        return ("", false)
+    end
+    return (clang_path, true)
+end
+
 function _hip_runtime_version()
     v_ref = Ref{Cint}()
     res = ccall((:hipRuntimeGetVersion, libhip), UInt32, (Ptr{Cint},), v_ref)
@@ -42,6 +52,8 @@ global rel_libdir::String = Sys.islinux() ? "" : "bin"
 global libhsaruntime::String = ""
 global lld_path::String = ""
 global lld_artifact::Bool = false
+global clang_path::String = ""
+global clang_artifact::Bool = false
 global libhip::String = ""
 global libdevice_libs::String = ""
 global librocblas::String = ""
@@ -78,6 +90,11 @@ function __init__()
         lld_path, lld_artifact = get_ld_lld(rocm_path)
         global lld_path = lld_path
         global lld_artifact = lld_artifact
+
+        # Clang
+        clang_path, clang_artifact = get_clang(rocm_path)
+        global clang_path = clang_path
+        global clang_artifact = clang_artifact
         global libhip = find_rocm_library(Sys.islinux() ? "libamdhip64" : "amdhip64"; rocm_path)
 
         global libdevice_libs = find_device_libs(rocm_path)

src/discovery/utils.jl

@@ -162,3 +162,20 @@ function find_ld_lld(rocm_path::String)::String
     end
     return ""
 end
+
+function find_clang(rocm_path::String)::String
+    clang_name = "clang" * (Sys.iswindows() ? ".exe" : "")
+
+    dirs = (joinpath(rocm_path, "llvm", "bin"), joinpath(rocm_path, "bin"))
+    hipconfig = Sys.which("hipconfig")
+    if !isnothing(hipconfig)
+        clang_path = read(`$hipconfig --hipclangpath`, String)
+        dirs = (dirs..., clang_path)
+    end
+    for dir in dirs
+        exp_clang_path = joinpath(dir, clang_name)
+        ispath(exp_clang_path) || continue
+        return exp_clang_path
+    end
+    return ""
+end