#!/bin/bash

set -e
 
OPENCL_ROOT=@OPENCL_ROOT@
BITCODE_DIR=${OPENCL_ROOT}/lib/x86_64/bitcode
CLANG=${OPENCL_ROOT}/bin/x86_64/clang
LLVM_LINK=${OPENCL_ROOT}/bin/x86_64/llvm-link

TRIPLE=@TRIPLE@

if [[ "$AMD_DEBUG_AMDGIZ" == 1 ]]; then
  TRIPLE=amdgcn-amd-amdhsa-amdgizcl
fi

gfxip=803
 
while (( "$#" ))
do
    arg="$1"
    case "$arg" in
    -o)
        shift
        output_file=$1
        ;;
    -mcpu=gfx*)
        gfxip=${1##*gfx}
        ;&
    -*)
        options="${options} $1"
        ;;
    *)
        input_file=$1
        ;;
    esac
    shift
done
 
 
${CLANG} -c -emit-llvm \
-target $TRIPLE -x cl \
-D__AMD__=1  \
-D__gfx${gfxip}__=1  \
-D__gfx${gfxip}=1  \
-D__OPENCL_VERSION__=120  \
-D__IMAGE_SUPPORT__=1 \
-O3 \
-m64 \
-cl-kernel-arg-info \
-cl-std=CL1.2 \
-mllvm -amdgpu-early-inline-all \
-Xclang -target-feature -Xclang -code-object-v3 \
-Xclang -cl-ext=+cl_khr_fp64,+cl_khr_global_int32_base_atomics,+cl_khr_global_int32_extended_atomics,+cl_khr_local_int32_base_atomics,+cl_khr_local_int32_extended_atomics,+cl_khr_int64_base_atomics,+cl_khr_int64_extended_atomics,+cl_khr_3d_image_writes,+cl_khr_byte_addressable_store,+cl_khr_gl_sharing,+cl_amd_media_ops,+cl_amd_media_ops2,+cl_khr_subgroups \
-include ${OPENCL_ROOT}/include/opencl-c.h \
${options} -o ${output_file}.orig.bc ${input_file}
 
${LLVM_LINK} -f -o ${output_file}.linked.bc ${output_file}.orig.bc \
$BITCODE_DIR/opencl.amdgcn.bc \
$BITCODE_DIR/ocml.amdgcn.bc \
$BITCODE_DIR/ockl.amdgcn.bc \
$BITCODE_DIR/irif.amdgcn.bc \
$BITCODE_DIR/oclc_correctly_rounded_sqrt_off.amdgcn.bc \
$BITCODE_DIR/oclc_daz_opt_on.amdgcn.bc \
$BITCODE_DIR/oclc_finite_only_off.amdgcn.bc \
$BITCODE_DIR/oclc_isa_version_${gfxip}.amdgcn.bc \
$BITCODE_DIR/oclc_unsafe_math_off.amdgcn.bc |& tee ${output_file}.linked.bc.out

# Fail on warnings
if grep -qi warning ${output_file}.linked.bc.out; then
    echo "llvm-link failed!"
    exit 1
fi
 
${CLANG} \
-target $TRIPLE \
-O3 \
-m64 \
-cl-kernel-arg-info \
-mllvm -amdgpu-internalize-symbols -mllvm -amdgpu-early-inline-all \
-Xclang -target-feature -Xclang -code-object-v3 \
${options} -o ${output_file} ${output_file}.linked.bc

# Remove extra files
rm ${output_file}.linked.bc*