Fix linking of AMD device libraries (#296)

It's weird that it fails without `-mno-link-builtin-bitcode-postopt`. I've tested it only on ROCm 6.2, might be broken on older or newer ROCm

2 files changed, 16 insertions, 13 deletions

diff --git a/comgr/src/lib.rs b/comgr/src/lib.rs
index 129dc14..94ba6ef 100644
--- a/comgr/src/lib.rs
+++ b/comgr/src/lib.rs
@@ -127,24 +127,27 @@ pub fn compile_bitcode(
         ptx_impl,

     )?;

     bitcode_data_set.add(&stdlib_bitcode_data)?;

-    let lang_action_info = ActionInfo::new()?;

-    lang_action_info.set_isa_name(gcn_arch)?;

-    lang_action_info.set_language(amd_comgr_language_t::AMD_COMGR_LANGUAGE_LLVM_IR)?;

-    let with_device_libs = do_action(

-        &bitcode_data_set,

-        &lang_action_info,

-        amd_comgr_action_kind_t::AMD_COMGR_ACTION_COMPILE_SOURCE_WITH_DEVICE_LIBS_TO_BC,

-    )?;

+    let linking_info = ActionInfo::new()?;

     let linked_data_set = do_action(

-        &with_device_libs,

-        &lang_action_info,

+        &bitcode_data_set,

+        &linking_info,

         amd_comgr_action_kind_t::AMD_COMGR_ACTION_LINK_BC_TO_BC,

     )?;

+    let link_with_device_libs_info = ActionInfo::new()?;

+    link_with_device_libs_info.set_isa_name(gcn_arch)?;

+    link_with_device_libs_info.set_language(amd_comgr_language_t::AMD_COMGR_LANGUAGE_LLVM_IR)?;

+    // This makes no sense, but it makes ockl linking work

+    link_with_device_libs_info.set_options([c"-Xclang", c"-mno-link-builtin-bitcode-postopt"].into_iter())?;

+    let with_device_libs = do_action(

+        &linked_data_set,

+        &link_with_device_libs_info,

+        amd_comgr_action_kind_t::AMD_COMGR_ACTION_COMPILE_SOURCE_WITH_DEVICE_LIBS_TO_BC,

+    )?;

     let compile_action_info = ActionInfo::new()?;

     compile_action_info.set_isa_name(gcn_arch)?;

     compile_action_info.set_options(iter::once(c"-O3"))?;

     let reloc_data_set = do_action(

-        &linked_data_set,

+        &with_device_libs,

         &compile_action_info,

         amd_comgr_action_kind_t::AMD_COMGR_ACTION_CODEGEN_BC_TO_RELOCATABLE,

     )?;

diff --git a/ptx/lib/zluda_ptx_impl.cpp b/ptx/lib/zluda_ptx_impl.cpp
index f1b416d..d0ec853 100644
--- a/ptx/lib/zluda_ptx_impl.cpp
+++ b/ptx/lib/zluda_ptx_impl.cpp
@@ -1,5 +1,5 @@
-// Every time this file changes it must te rebuilt, you need llvm-17:

-//  /opt/rocm/llvm/bin/clang -Wall -Wextra -Wsign-compare -Wconversion -x hip zluda_ptx_impl.cpp -nogpulib -O3 -mno-wavefrontsize64 -o zluda_ptx_impl.bc -emit-llvm -c --offload-device-only --offload-arch=gfx1010 && llvm-dis-17 zluda_ptx_impl.bc -o - | sed '/@llvm.used/d' | sed '/wchar_size/d' | sed '/llvm.module.flags/d' | sed 's/define hidden/define linkonce_odr/g' | sed 's/\"target-cpu\"=\"gfx1010\"//g' | sed -E 's/\"target-features\"=\"[^\"]+\"//g' | llvm-as-17 - -o  zluda_ptx_impl.bc && llvm-dis-17 zluda_ptx_impl.bc

+// Every time this file changes it must te rebuilt, you need `rocm-llvm-dev` and `llvm-17`:

+//  /opt/rocm/llvm/bin/clang -Wall -Wextra -Wsign-compare -Wconversion -x hip zluda_ptx_impl.cpp -nogpulib -O3 -mno-wavefrontsize64 -o zluda_ptx_impl.bc -emit-llvm -c --offload-device-only --offload-arch=gfx1010 && /opt/rocm/llvm/bin/llvm-dis zluda_ptx_impl.bc -o - | sed '/@llvm.used/d' | sed '/wchar_size/d' | sed '/llvm.module.flags/d' | sed 's/define hidden/define linkonce_odr/g' | sed 's/\"target-cpu\"=\"gfx1010\"//g' | sed -E 's/\"target-features\"=\"[^\"]+\"//g' | sed 's/ nneg / /g' | sed 's/ disjoint / /g' | llvm-as-17 - -o  zluda_ptx_impl.bc && /opt/rocm/llvm/bin/llvm-dis zluda_ptx_impl.bc

 

 #include <cstddef>

 #include <cstdint>