[sycl-post-link][Driver] Pass in -emit-only-kernels-as-entry-point option by default (#12060)

sycl-post-link is one of the tools invoked during compilation of SYCL device code. One of the options passed to this tool is ‘emit-only-kernels-as-entry-points’. This option was introduced to stop treating `SYCL_EXTERNAL` functions as entry points. However, it was turned on by default only for FPGAs and turned off for other backends. We are noticing that there are no tangible benefits to treating `SYCL_EXTERNAL` functions as entry points and that treating them as entry points results in unnecessary device code bloating. This is a pain point reported by several customers. We propose to turn on `
emit-only-kernels-as-entry-points` by default for all Intel backends. However, we want to expose this option to users and allow them to turn it off if required. Hence we propose the following command line option
to be added into compiler:  -f[no]-sycl-remove-unused-external-funcs // Linux and Windows
This option takes no arguments

Thanks
---------

Signed-off-by: Sudarsanam, Arvind <[email protected]>

Author: asudarsa

clang/include/clang/Driver/Options.td

@@ -3943,6 +3943,10 @@ def foperator_arrow_depth_EQ : Joined<["-"], "foperator-arrow-depth=">, Group<f_
   Visibility<[ClangOption, CC1Option]>,
   HelpText<"Maximum number of 'operator->'s to call for a member access">,
   MarshallingInfoInt<LangOpts<"ArrowDepth">, "256">;
+def fsycl_remove_unused_external_funcs : Flag<["-"], "fsycl-remove-unused-external-funcs">,
+  Group<sycl_Group>, HelpText<"Allow removal of unused `SYCL_EXTERNAL` functions (default)">;
+def fno_sycl_remove_unused_external_funcs : Flag<["-"], "fno-sycl-remove-unused-external-funcs">,
+  Group<sycl_Group>, HelpText<"Prevent removal of unused `SYCL_EXTERNAL` functions">;
 
 def fsave_optimization_record : Flag<["-"], "fsave-optimization-record">,
   Visibility<[ClangOption, FlangOption]>,

clang/lib/Driver/ToolChains/Clang.cpp

@@ -10215,10 +10215,13 @@ void SYCLPostLink::ConstructJob(Compilation &C, const JobAction &JA,
     addArgs(CmdArgs, TCArgs, {"-split=auto"});
   }
 
-  // On FPGA target 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.
-  if (T.getArchName() == "spir64_fpga")
+  // 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) &&
+      !T.isNVPTX() && !T.isAMDGPU())
     addArgs(CmdArgs, TCArgs, {"-emit-only-kernels-as-entry-points"});
 
   // OPT_fsycl_device_code_split is not checked as it is an alias to

clang/test/Driver/sycl-offload-intelfpga.cpp

@@ -41,17 +41,6 @@
 // RUN:   | FileCheck -check-prefix=CHK-NON-KERNEL-ENTRY-POINTS %s
 // CHK-NON-KERNEL-ENTRY-POINTS: sycl-post-link{{.*}} "-emit-only-kernels-as-entry-points"
 
-/// Non-FPGA targets should not imply -emit-only-kernels-as-entry-points in sycl-post-link
-// RUN:   %clang -### -fsycl -fsycl-targets=spir64_fpga,spir64_gen %s 2>&1 \
-// RUN:   | FileCheck %s --check-prefix=CHK-NON-KERNEL-ENTRY-POINTS-NEG-1
-// CHK-NON-KERNEL-ENTRY-POINTS-NEG-1: clang{{.*}} "-triple" "spir64_fpga-unknown-unknown"
-// CHK-NON-KERNEL-ENTRY-POINTS-NEG-1: sycl-post-link{{.*}} "-emit-only-kernels-as-entry-points"
-// CHK-NON-KERNEL-ENTRY-POINTS-NEG-1: clang{{.*}} "-triple" "spir64_gen-unknown-unknown"
-// CHK-NON-KERNEL-ENTRY-POINTS-NEG-1: sycl-post-link
-// CHK-NON-KERNEL-ENTRY-POINTS-NEG-1-NOT: "-emit-only-kernels-as-entry-points"
-// RUN:   %clang -### -fsycl %s 2>&1 | FileCheck %s --check-prefix=CHK-NON-KERNEL-ENTRY-POINTS-NEG-2
-// CHK-NON-KERNEL-ENTRY-POINTS-NEG-2-NOT: "-emit-only-kernels-as-entry-points"
-
 /// -fsycl-disable-range-rounding is applied to all compilations if fpga is used
 // RUN:   %clangxx -### -target x86_64-unknown-linux-gnu -fsycl -fsycl-targets=spir64_fpga-unknown-unknown,spir64_gen-unknown-unknown %s 2>&1 \
 // RUN:   | FileCheck -check-prefix=CHK-RANGE-ROUNDING-MULTI %s

clang/test/Driver/sycl-offload.cpp

@@ -170,3 +170,11 @@
 /// Check if the clang with fsycl adds C++ libraries to the link line
 //  RUN:  %clang -### -target x86_64-unknown-linux-gnu -fsycl %s 2>&1 | FileCheck -check-prefix=CHECK-FSYCL-WITH-CLANG %s
 // CHECK-FSYCL-WITH-CLANG: "-lstdc++"
+
+/// Check selective passing of -emit-only-kernels-as-entry-points to sycl-post-link tool
+// RUN: %clang -### -target x86_64-unknown-linux-gnu -fsycl -fsycl-targets=spir64_fpga %s 2>&1 | FileCheck -check-prefix=CHECK_SYCL_POST_LINK_OPT_PASS %s
+// RUN: %clang -### -target x86_64-unknown-linux-gnu -fsycl -fsycl-targets=spir64_gen %s 2>&1 | FileCheck -check-prefix=CHECK_SYCL_POST_LINK_OPT_PASS %s
+// CHECK_SYCL_POST_LINK_OPT_PASS: sycl-post-link{{.*}}emit-only-kernels-as-entry-points
+// RUN: %clang -### -target x86_64-unknown-linux-gnu -fsycl -fsycl-targets=spir64_gen -fno-sycl-remove-unused-external-funcs %s 2>&1 | FileCheck -check-prefix=CHECK_SYCL_POST_LINK_OPT_NO_PASS %s
+// CHECK_SYCL_POST_LINK_OPT_NO_PASS-NOT: sycl-post-link{{.*}}emit-only-kernels-as-entry-points
+