[SYCL][NewOffloadModel] Support optional kernel features and fix ocloc argument parsing (#20470)

This patch adds support for optional kernel features and fixes ocloc
command parsing with the following changes:

(1) Enable optional kernel features for the new offloading model by
passing the device architecture name to the `sycl-post-link` command
constructed in `ClangLinkerWrapper.cpp`. This invokes `sycl-post-link`
with the device architecture when Intel GPU targets are specified in
`-fsycl-targets`. A new test has been added in `aot-gpu.cpp` to verify
this functionality.

(2) The original implementation in `ClangLinkerWrapper.cpp` incorrectly
constructed `ocloc` command line arguments by concatenating all
arguments into a single string, causing parsing failures in the
executor. This patch fixes the issue by properly splitting arguments on
whitespace boundaries and rejoining them into a correctly formatted
command string.

(3) Add test cases running with `-O0 -g` flag to `fp64-conv-emu-1.cpp`
and `fp64-conv-emu-2.cpp`

The following tests now pass (fail before) with the new offload model
after applying this patch:
1. fp64-conv-emu-1.cpp
2. fp64-conv-emu-2.cpp
3. AOT/double.cpp
4. AOT/half.cpp
5. AOT/reqd-sg-size.cpp
6. BFloat16/bfloat16_example_aot_gpu.cpp
7. DeviceArchitecture/device_architecture_comparison_on_device_aot.cpp
8. AOT/fallback.cpp

Author: YixingZhang007

clang/test/Driver/clang-linker-wrapper.cpp

@@ -63,13 +63,20 @@
 // CHK-CMDS-AOT-GEN: spirv-to-ir-wrapper{{.*}} -o [[FIRSTLLVMLINKIN:.*]].bc --llvm-spirv-opts --spirv-preserve-auxdata --spirv-target-env=SPV-IR --spirv-builtin-format=global
 // CHK-CMDS-AOT-GEN-NEXT: llvm-link{{.*}} --suppress-warnings [[FIRSTLLVMLINKIN]].bc -o [[FIRSTLLVMLINKOUT:.*]].bc
 // CHK-CMDS-AOT-GEN-NEXT: llvm-link{{.*}} -only-needed --suppress-warnings [[FIRSTLLVMLINKOUT]].bc {{.*}}.bc -o [[SECONDLLVMLINKOUT:.*]].bc
-// CHK-CMDS-AOT-GEN-NEXT: sycl-post-link{{.*}} SYCL_POST_LINK_OPTIONS -o [[SYCLPOSTLINKOUT:.*]].table [[SECONDLLVMLINKOUT]].bc
+// Check that target specified by -fsycl-targets is passed to sycl-post-link for filtering.
+// CHK-CMDS-AOT-GEN-NEXT: sycl-post-link{{.*}} SYCL_POST_LINK_OPTIONS -o intel_gpu_pvc,[[SYCLPOSTLINKOUT:.*]].table [[SECONDLLVMLINKOUT]].bc
 // CHK-CMDS-AOT-GEN-NEXT: llvm-spirv{{.*}} LLVM_SPIRV_OPTIONS -o {{.*}}
 // CHK-CMDS-AOT-GEN-NEXT: ocloc{{.*}} -output_no_suffix -spirv_input -device pvc{{.*}} -output {{.*}} -file {{.*}}
 // CHK-CMDS-AOT-GEN-NEXT: offload-wrapper: output: [[WRAPPEROUT:.*]].bc, input: {{.*}}, compile-opts: , link-opts:
 // CHK-CMDS-AOT-GEN-NEXT: clang{{.*}} -c -o [[LLCOUT:.*]].o [[WRAPPEROUT]].bc
 // CHK-CMDS-AOT-GEN-NEXT: "{{.*}}/ld" -- HOST_LINKER_FLAGS -dynamic-linker HOST_DYN_LIB -o a.out [[LLCOUT]].o HOST_LIB_PATH HOST_STAT_LIB {{.*}}.o
 
+// Check that when --gpu-tool-arg is specified in clang-linker-wrapper 
+// (happen when AOT device is specified via -Xsycl-target-backend '-device pvc' in clang), 
+// the target is not passed to sycl-post-link for filtering.
+// RUN: clang-linker-wrapper -sycl-embed-ir -sycl-device-libraries=%t1.devicelib.o -sycl-post-link-options="SYCL_POST_LINK_OPTIONS" -llvm-spirv-options="LLVM_SPIRV_OPTIONS" "--host-triple=x86_64-unknown-linux-gnu" "--gpu-tool-arg=-device pvc" "--linker-path=/usr/bin/ld" "--" HOST_LINKER_FLAGS "-dynamic-linker" HOST_DYN_LIB "-o" "a.out" HOST_LIB_PATH HOST_STAT_LIB %t1.o --dry-run 2>&1 | FileCheck -check-prefix=CHK-NO-CMDS-AOT-GEN %s
+// CHK-NO-CMDS-AOT-GEN-NOT: sycl-post-link{{.*}} -o intel_gpu_pv,{{.*}}
+
 /// Check for list of commands for standalone clang-linker-wrapper run for sycl (AOT for Intel CPU)
 // -------
 // Generate .o file as linker wrapper input.

clang/tools/clang-linker-wrapper/ClangLinkerWrapper.cpp

@@ -727,10 +727,24 @@ runSYCLPostLinkTool(ArrayRef<StringRef> InputFiles, const ArgList &Args) {
       createOutputFile(sys::path::filename(ExecutableName), "table");
   if (!TempFileOrErr)
     return TempFileOrErr.takeError();
+  std::string OutputPathWithArch = TempFileOrErr->str();
+
+  // Enable the driver to invoke sycl-post-link with the device architecture
+  // when Intel GPU targets are passed in -fsycl-targets.
+  // OPT_gpu_tool_arg_EQ is checked to ensure the device architecture is not
+  // passed through -Xsycl-target-backend=spir64_gen "-device <arch>" format
+  const llvm::Triple Triple(Args.getLastArgValue(OPT_triple_EQ));
+  StringRef Arch = Args.getLastArgValue(OPT_arch_EQ);
+  StringRef IsGPUTool = Args.getLastArgValue(OPT_gpu_tool_arg_EQ);
+
+  if (Triple.getSubArch() == llvm::Triple::SPIRSubArch_gen && !Arch.empty() &&
+      IsGPUTool.empty() && Arch != "*")
+    OutputPathWithArch = "intel_gpu_" + Arch.str() + "," + OutputPathWithArch;
+  else if (Triple.getSubArch() == llvm::Triple::SPIRSubArch_x86_64)
+    OutputPathWithArch = "spir64_x86_64," + OutputPathWithArch;
 
   SmallVector<StringRef, 8> CmdArgs;
   CmdArgs.push_back(*SYCLPostLinkPath);
-  const llvm::Triple Triple(Args.getLastArgValue(OPT_triple_EQ));
   Arg *SYCLDeviceLibLoc = Args.getLastArg(OPT_sycl_device_library_location_EQ);
   if (SYCLDeviceLibLoc && !Triple.isSPIRAOT()) {
     std::string SYCLDeviceLibLocParam = SYCLDeviceLibLoc->getValue();
@@ -748,7 +762,7 @@ runSYCLPostLinkTool(ArrayRef<StringRef> InputFiles, const ArgList &Args) {
   SYCLPostLinkOptions.split(CmdArgs, " ", /* MaxSplit = */ -1,
                             /* KeepEmpty = */ false);
   CmdArgs.push_back("-o");
-  CmdArgs.push_back(*TempFileOrErr);
+  CmdArgs.push_back(Args.MakeArgString(OutputPathWithArch));
   for (auto &File : InputFiles)
     CmdArgs.push_back(File);
   if (Error Err = executeCommands(*SYCLPostLinkPath, CmdArgs))

sycl/test-e2e/OptionalKernelFeatures/fp64-conv-emu-1.cpp

@@ -20,6 +20,9 @@
 // RUN: %clangxx -fsycl -fsycl-targets=intel_gpu_dg2_g10,intel_gpu_dg2_g11,intel_gpu_dg2_g12,intel_gpu_pvc,intel_gpu_mtl_h,intel_gpu_mtl_u -fsycl-fp64-conv-emu %O0 %s -o %t.out
 // RUN: %{run} %t.out
 
+// RUN: %clangxx -fsycl -fsycl-targets=intel_gpu_dg2_g10,intel_gpu_dg2_g11,intel_gpu_dg2_g12,intel_gpu_pvc,intel_gpu_mtl_h,intel_gpu_mtl_u -fsycl-fp64-conv-emu -g %O0 %s -o %t.debug.out
+// RUN: %{run} %t.debug.out
+
 // Tests that aspect::fp64 is not emitted correctly when -fsycl-fp64-conv-emu
 // flag is used.
 

sycl/test-e2e/OptionalKernelFeatures/fp64-conv-emu-2.cpp

@@ -23,6 +23,9 @@
 // RUN: %clangxx -fsycl -fsycl-targets=intel_gpu_dg2_g10,intel_gpu_dg2_g11,intel_gpu_dg2_g12,intel_gpu_pvc,intel_gpu_mtl_h,intel_gpu_mtl_u -fsycl-fp64-conv-emu %O0 %s -o %t.out
 // RUN: %{run} %t.out
 
+// RUN: %clangxx -fsycl -fsycl-targets=intel_gpu_dg2_g10,intel_gpu_dg2_g11,intel_gpu_dg2_g12,intel_gpu_pvc,intel_gpu_mtl_h,intel_gpu_mtl_u -fsycl-fp64-conv-emu -g %O0 %s -o %t.debug.out
+// RUN: %{run} %t.debug.out
+
 #include <sycl/detail/core.hpp>
 using namespace sycl;