[clang] d1d80cc - [HIP] Make the new driver bundle outputs for device-only (#84534)

[via cfe-commits]

Author: Joseph Huber
Date: 2024-03-11T14:09:59-05:00
New Revision: d1d80cc3197faa4194cddcc79ff704b7d4c5b9e4

URL: https://github.com/llvm/llvm-project/commit/d1d80cc3197faa4194cddcc79ff704b7d4c5b9e4
DIFF: https://github.com/llvm/llvm-project/commit/d1d80cc3197faa4194cddcc79ff704b7d4c5b9e4.diff

LOG: [HIP] Make the new driver bundle outputs for device-only (#84534)

Summary:
The current behavior of HIP is that when --offload-device-only is set it
still bundles the outputs into a fat binary. Even though this is
different from how all the other targets handle this, it seems to be
dependned on by some tooling so just make it backwards compatible for
the `-fno-gpu-rdc` case.

Added: 
    

Modified: 
    clang/lib/Driver/Driver.cpp
    clang/test/Driver/hip-binding.hip

Removed: 
    


################################################################################
diff  --git a/clang/lib/Driver/Driver.cpp b/clang/lib/Driver/Driver.cpp
index fce43430a91374..190782a79a2456 100644
--- a/clang/lib/Driver/Driver.cpp
+++ b/clang/lib/Driver/Driver.cpp
@@ -4638,7 +4638,12 @@ Action *Driver::BuildOffloadingActions(Compilation &C,
     }
   }
 
-  if (offloadDeviceOnly())
+  // All kinds exit now in device-only mode except for non-RDC mode HIP.
+  if (offloadDeviceOnly() &&
+      (!C.isOffloadingHostKind(Action::OFK_HIP) ||
+       !Args.hasFlag(options::OPT_gpu_bundle_output,
+                     options::OPT_no_gpu_bundle_output, true) ||
+       Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, false)))
     return C.MakeAction<OffloadAction>(DDeps, types::TY_Nothing);
 
   if (OffloadActions.empty())
@@ -4671,6 +4676,10 @@ Action *Driver::BuildOffloadingActions(Compilation &C,
              nullptr, C.getActiveOffloadKinds());
   }
 
+  // HIP wants '--offload-device-only' to create a fatbinary by default.
+  if (offloadDeviceOnly())
+    return C.MakeAction<OffloadAction>(DDep, types::TY_Nothing);
+
   // If we are unable to embed a single device output into the host, we need to
   // add each device output as a host dependency to ensure they are still built.
   bool SingleDeviceOutput = !llvm::any_of(OffloadActions, [](Action *A) {

diff  --git a/clang/test/Driver/hip-binding.hip b/clang/test/Driver/hip-binding.hip
index 79ec2039edb74c..c116ad80a8ad80 100644
--- a/clang/test/Driver/hip-binding.hip
+++ b/clang/test/Driver/hip-binding.hip
@@ -65,9 +65,18 @@
 // MULTI-D-ONLY-NEXT: # "amdgcn-amd-amdhsa" - "AMDGCN::Linker", inputs: ["[[GFX90a]]"], output: "[[GFX90a_OUT:.+]]"
 //
 // RUN: not %clang -### --target=x86_64-linux-gnu --offload-new-driver -ccc-print-bindings -nogpulib -nogpuinc \
-// RUN:        --offload-arch=gfx90a --offload-arch=gfx908 --offload-device-only -c -o %t %s 2>&1 \
+// RUN:        --no-gpu-bundle-output --offload-arch=gfx90a --offload-arch=gfx908 --offload-device-only -c -o %t %s 2>&1 \
+// RUN: | FileCheck -check-prefix=MULTI-D-ONLY-NO-BUNDLE-O %s
+// MULTI-D-ONLY-NO-BUNDLE-O: error: cannot specify -o when generating multiple output files
+
+// RUN: %clang -### --target=x86_64-linux-gnu --offload-new-driver -ccc-print-bindings -nogpulib -nogpuinc \
+// RUN:        --gpu-bundle-output --offload-arch=gfx90a --offload-arch=gfx908 --offload-device-only -c -o a.out %s 2>&1 \
 // RUN: | FileCheck -check-prefix=MULTI-D-ONLY-O %s
-// MULTI-D-ONLY-O: error: cannot specify -o when generating multiple output files
+//      MULTI-D-ONLY-O: "amdgcn-amd-amdhsa" - "clang", inputs: ["[[INPUT:.+]]"], output: "[[GFX908_OBJ:.+]]"
+// MULTI-D-ONLY-O-NEXT: "amdgcn-amd-amdhsa" - "AMDGCN::Linker", inputs: ["[[GFX908_OBJ]]"], output: "[[GFX908:.+]]"
+// MULTI-D-ONLY-O-NEXT: "amdgcn-amd-amdhsa" - "clang", inputs: ["[[INPUT]]"], output: "[[GFX90A_OBJ:.+]]"
+// MULTI-D-ONLY-O-NEXT: "amdgcn-amd-amdhsa" - "AMDGCN::Linker", inputs: ["[[GFX90A_OBJ]]"], output: "[[GFX90A:.+]]"
+// MULTI-D-ONLY-O-NEXT: "amdgcn-amd-amdhsa" - "AMDGCN::Linker", inputs: ["[[GFX908]]", "[[GFX90A]]"], output: "a.out"
 
 //
 // Check to ensure that we can use '-fsyntax-only' for HIP output with the new