[SYCL][Driver][thinLTO] Don't pass -emit-only-kernels-as-entry-points to sycl-post-link in early splitting (#14991)

In early splitting with thinLTO, we could be generating an `.o` to be
linked in with other files later. Right now, passing
`-emit-only-kernels-as-entry-points` to `sycl-post-link` can cause
functions to get dropped even though they are used by some other `.o`.

We will need to prune non-entry points inside the thinLTO processing
inside `clang-linker-wrapper`, but that's not implemented yet.

---------

Signed-off-by: Sarnie, Nick <nick.sarnie@intel.com>

Author: sarnex

clang/lib/Driver/ToolChains/Clang.cpp

@@ -10741,6 +10741,32 @@ static void getNonTripleBasedSYCLPostLinkOpts(const ToolChain &TC,
     addArgs(PostLinkArgs, TCArgs, {"-support-dynamic-linking"});
 }
 
+// On Intel targets we don't need non-kernel functions as entry points,
+// because it only increases amount of code for device compiler to handle,
+// without any actual benefits.
+// TODO: Try to extend this feature for non-Intel GPUs.
+static bool shouldEmitOnlyKernelsAsEntryPoints(const ToolChain &TC,
+                                               const llvm::opt::ArgList &TCArgs,
+                                               llvm::Triple Triple) {
+  if (TCArgs.hasFlag(options::OPT_fno_sycl_remove_unused_external_funcs,
+                     options::OPT_fsycl_remove_unused_external_funcs, false))
+    return false;
+  if (isSYCLNativeCPU(TC))
+    return false;
+  // When supporting dynamic linking, non-kernels in a device image can be
+  // called.
+  if (supportDynamicLinking(TCArgs))
+    return false;
+  if (Triple.isNVPTX() || Triple.isAMDGPU())
+    return false;
+  bool IsUsingLTO = TC.getDriver().isUsingLTO(/*IsDeviceOffloadAction=*/true);
+  auto LTOMode = TC.getDriver().getLTOMode(/*IsDeviceOffloadAction=*/true);
+  // With thinLTO, final entry point handing is done in clang-linker-wrapper
+  if (IsUsingLTO && LTOMode == LTOK_Thin)
+    return false;
+  return true;
+}
+
 // Add any sycl-post-link options that rely on a specific Triple in addition
 // to user supplied options. This function is invoked only for the old
 // offloading model. For the new offloading model, a slightly modified version
@@ -10778,17 +10804,7 @@ static void getTripleBasedSYCLPostLinkOpts(const ToolChain &TC,
       (Triple.getArchName() != "spir64_fpga"))
     addArgs(PostLinkArgs, TCArgs, {"-split=auto"});
 
-  // On Intel targets we don't need non-kernel functions as entry points,
-  // because it only increases amount of code for device compiler to handle,
-  // without any actual benefits.
-  // TODO: Try to extend this feature for non-Intel GPUs.
-  if ((!TCArgs.hasFlag(options::OPT_fno_sycl_remove_unused_external_funcs,
-                       options::OPT_fsycl_remove_unused_external_funcs,
-                       false) &&
-       !isSYCLNativeCPU(TC)) &&
-      // When supporting dynamic linking, non-kernels in a device image can be
-      // called.
-      !supportDynamicLinking(TCArgs) && !Triple.isNVPTX() && !Triple.isAMDGPU())
+  if (shouldEmitOnlyKernelsAsEntryPoints(TC, TCArgs, Triple))
     addArgs(PostLinkArgs, TCArgs, {"-emit-only-kernels-as-entry-points"});
 
   if (!Triple.isAMDGCN())

clang/test/Driver/sycl-lto.cpp

@@ -9,9 +9,10 @@
 // CHECK_SPLIT_ERROR: '-fsycl-device-code-split=off' is not supported when '-foffload-lto=thin' is set with '-fsycl'
 
 // Verify there's no error and we see the expected cc1 flags and tool invocations with the new offload driver.
-// RUN: %clangxx -fsycl --offload-new-driver -foffload-lto=thin %s -### 2>&1 | FileCheck -check-prefix=CHECK_SUPPORTED %s
+// RUN: %clangxx -fsycl --offload-new-driver -foffload-lto=thin %s -### 2>&1 | \
+// RUN: FileCheck -check-prefix=CHECK_SUPPORTED -implicit-check-not=-emit-only-kernels-as-entry-points %s
 // CHECK_SUPPORTED: clang{{.*}} "-cc1" "-triple" "spir64-unknown-unknown" {{.*}} "-flto=thin" "-flto-unit"
-// CHECK_SUPPORTED: sycl-post-link{{.*}}
+// CHECK_SUPPORTED: sycl-post-link
 // CHECK_SUPPORTED-NOT: -properties
 // CHECK_SUPPORTED-NEXT: file-table-tform{{.*}}
 // CHECK_SUPPORTED-NEXT: llvm-foreach{{.*}} "--" {{.*}}clang{{.*}} "-fsycl-is-device"{{.*}} "-flto=thin" "-flto-unit"