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

[Joseph Huber via cfe-commits]

https://github.com/jhuber6 created https://github.com/llvm/llvm-project/pull/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.


>From 99a769ec7ffaa7728847fdf2f67a1be11ce98f2b Mon Sep 17 00:00:00 2001
From: Joseph Huber <huberjn at outlook.com>
Date: Fri, 8 Mar 2024 12:49:38 -0600
Subject: [PATCH] [HIP] Make the new driver bundle outputs for device-only

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.
---
 clang/lib/Driver/Driver.cpp       | 10 +++++++++-
 clang/test/Driver/hip-binding.hip | 10 +++++++---
 2 files changed, 16 insertions(+), 4 deletions(-)

diff --git a/clang/lib/Driver/Driver.cpp b/clang/lib/Driver/Driver.cpp
index fce43430a91374..eba43d97431364 100644
--- a/clang/lib/Driver/Driver.cpp
+++ b/clang/lib/Driver/Driver.cpp
@@ -4638,7 +4638,10 @@ 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_fgpu_rdc, options::OPT_fno_gpu_rdc, false)))
     return C.MakeAction<OffloadAction>(DDeps, types::TY_Nothing);
 
   if (OffloadActions.empty())
@@ -4671,6 +4674,11 @@ Action *Driver::BuildOffloadingActions(Compilation &C,
              nullptr, C.getActiveOffloadKinds());
   }
 
+  // HIP wants '--offload-device-only' to create a fatbinary by default.
+  if (offloadDeviceOnly() && C.isOffloadingHostKind(Action::OFK_HIP) &&
+      !Args.hasFlag(options::OPT_fgpu_rdc, options::OPT_fno_gpu_rdc, false))
+    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..cb17112c28d90a 100644
--- a/clang/test/Driver/hip-binding.hip
+++ b/clang/test/Driver/hip-binding.hip
@@ -64,10 +64,14 @@
 // MULTI-D-ONLY-NEXT: # "amdgcn-amd-amdhsa" - "clang", inputs: ["[[INPUT]]"], output: "[[GFX90a:.+]]"
 // 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: %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 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