[Driver][SYCL] Enable early AOT abilities when creating objects

clang/include/clang/Driver/Action.h

@@ -629,11 +629,20 @@ class OffloadUnbundlingJobAction final : public JobAction {
           DependentOffloadKind(DependentOffloadKind) {}
   };
 
+  /// Allow for a complete override of the target to unbundle.
+  /// This is used for specific unbundles used for SYCL AOT when generating full
+  /// device files that are bundled with the host object.
+  void setTargetString(std::string Target) { TargetString = Target; }
+
+  std::string getTargetString() const { return TargetString; }
+
 private:
   /// Container that keeps information about each dependence of this unbundling
   /// action.
   SmallVector<DependentActionInfo, 6> DependentActionInfoArray;
 
+  std::string TargetString;
+
 public:
   // Offloading unbundling doesn't change the type of output.
   OffloadUnbundlingJobAction(Action *Input);

clang/include/clang/Driver/Options.td

@@ -3858,7 +3858,14 @@ def ftarget_register_alloc_mode_EQ : Joined<["-"], "ftarget-register-alloc-mode=
   HelpText<"Specify a register allocation mode for specific hardware for use by supported "
   "target backends.">;
 def : Flag<["-"], "fsycl-rdc">, Visibility<[ClangOption, CLOption, DXCOption]>, Alias<fgpu_rdc>;
-def : Flag<["-"], "fno-sycl-rdc">, Visibility<[ClangOption, CLOption, DXCOption]>, Alias<fno_gpu_rdc>;
+def : Flag<["-"], "fno-sycl-rdc">,
+  Visibility<[ClangOption, CLOption, DXCOption]>, Alias<fno_gpu_rdc>,
+  HelpText<"Generate relocatable device code during SYCL offload target "
+           "compilation. Use of ‘-fno-sycl-rdc’ in combination with ‘-c’ will "
+           "produce final device binaries within the generated fat object. "
+           "When using this option, each kernel must be self-contained within "
+           "its translation unit (source file).  Therefore, the use of "
+           "SYCL_EXTERNAL is disallowed when this option is enabled.">;
 def fsycl_optimize_non_user_code : Flag<["-"], "fsycl-optimize-non-user-code">,
   Visibility<[ClangOption, CLOption, DXCOption, CC1Option]>,
   MarshallingInfoFlag<CodeGenOpts<"OptimizeSYCLFramework">>,

clang/lib/Driver/ToolChains/Clang.cpp

@@ -9216,7 +9216,24 @@ void OffloadBundler::ConstructJob(Compilation &C, const JobAction &JA,
                    ? Action::GetOffloadKindName(Action::OFK_SYCL)
                    : Action::GetOffloadKindName(CurKind);
     Triples += '-';
-    Triples += CurTC->getTriple().normalize();
+    // Incoming DeviceArch is set, break down the Current triple and add the
+    // device arch value to it.
+    // This is done for AOT targets only.
+    std::string DeviceArch;
+    llvm::Triple TargetTriple(CurTC->getTriple());
+    if (CurKind == Action::OFK_SYCL && TargetTriple.isSPIRAOT() &&
+        tools::SYCL::shouldDoPerObjectFileLinking(C))
+      DeviceArch = std::string("image");
+    if (CurKind != Action::OFK_Host && !DeviceArch.empty()) {
+      llvm::Triple T(CurTC->getTriple());
+      SmallString<128> ArchName(CurTC->getArchName());
+      ArchName += "_";
+      ArchName += DeviceArch.data();
+      T.setArchName(ArchName);
+      Triples += T.normalize();
+    } else {
+      Triples += CurTC->getTriple().normalize();
+    }
     if ((CurKind == Action::OFK_HIP || CurKind == Action::OFK_OpenMP ||
          CurKind == Action::OFK_Cuda || CurKind == Action::OFK_SYCL) &&
         !StringRef(CurDep->getOffloadingArch()).empty() &&
@@ -9467,7 +9484,16 @@ void OffloadBundler::ConstructJobMultipleOutputs(
     Triples += '-';
     Triples += types::getTypeName(types::TY_FPGA_Dependencies);
   }
-  CmdArgs.push_back(TCArgs.MakeArgString(Triples));
+  std::string TargetString(UA.getTargetString());
+  if (!TargetString.empty()) {
+    // The target string was provided, we will override the defaults and use
+    // the string provided.
+    SmallString<128> TSTriple("-targets=");
+    TSTriple += TargetString;
+    CmdArgs.push_back(TCArgs.MakeArgString(TSTriple));
+  } else {
+    CmdArgs.push_back(TCArgs.MakeArgString(Triples));
+  }
 
   // Get bundled file command.
   CmdArgs.push_back(

clang/lib/Driver/ToolChains/Gnu.cpp

@@ -597,7 +597,8 @@ void tools::gnutools::Linker::ConstructJob(Compilation &C, const JobAction &JA,
   // linked archives.  The unbundled information is a list of files and not
   // an actual object/archive.  Take that list and pass those to the linker
   // instead of the original object.
-  if (JA.isDeviceOffloading(Action::OFK_OpenMP)) {
+  if (JA.isDeviceOffloading(Action::OFK_OpenMP) ||
+      JA.isOffloading(Action::OFK_SYCL)) {
     InputInfoList UpdatedInputs;
     // Go through the Inputs to the link.  When a listfile is encountered, we
     // know it is an unbundled generated list.

clang/test/Driver/sycl-early-device-link.cpp

@@ -0,0 +1,131 @@
+// Testing for early AOT device linking.  These tests use -fno-sycl-rdc
+// -c to create final device binaries during the link step when using -fsycl.
+// Behavior is restricted to spir64_gen targets for now.
+
+// Create object that contains final device image
+// RUN: %clangxx -c -fno-sycl-rdc -fsycl -fsycl-targets=spir64_gen \
+// RUN:          --target=x86_64-unknown-linux-gnu -Xsycl-target-backend \
+// RUN:          "-device skl" --sysroot=%S/Inputs/SYCL -### %s 2>&1 \
+// RUN:  | FileCheck %s -check-prefix=CREATE_IMAGE
+// CREATE_IMAGE: clang{{.*}} "-triple" "spir64_gen-unknown-unknown"{{.*}} "-fsycl-is-device"{{.*}} "-o" "[[DEVICE_BC:.+\.bc]]"
+// CREATE_IMAGE: llvm-link{{.*}} "-o" "[[LIB_DEVICE_BC:.+\.bc]]"
+// CREATE_IMAGE: llvm-link{{.*}} "[[DEVICE_BC]]" "[[LIB_DEVICE_BC]]"{{.*}} "-o" "[[FINAL_DEVICE_BC:.+\.bc]]"
+// CREATE_IMAGE: sycl-post-link{{.*}} "-o" "[[POSTLINK_TABLE:.+\.table]]" "[[FINAL_DEVICE_BC]]"
+// CREATE_IMAGE: file-table-tform{{.*}} "-o" "[[TFORM_TXT:.+\.txt]]" "[[POSTLINK_TABLE]]"
+// CREATE_IMAGE: llvm-spirv{{.*}} "-o" "[[LLVMSPIRV_TXT:.+\.txt]]"{{.*}} "[[TFORM_TXT]]"
+// CREATE_IMAGE: ocloc{{.*}} "-output" "[[OCLOC_OUT:.+\.out]]" "-file" "[[LLVMSPIRV_TXT]]"{{.*}} "-device" "skl"
+// CREATE_IMAGE: file-table-tform{{.*}} "-o" "[[TFORM_TABLE:.+\.table]]" "[[POSTLINK_TABLE]]" "[[OCLOC_OUT]]"
+// CREATE_IMAGE: clang-offload-wrapper{{.*}} "-o=[[WRAPPER_BC:.+\.bc]]"
+// CREATE_IMAGE: llc{{.*}} "-o" "[[DEVICE_OBJECT:.+\.o]]" "[[WRAPPER_BC]]"
+// CREATE_IMAGE: append-file{{.*}} "--output=[[APPEND_SOURCE:.+\.cpp]]
+// CREATE_IMAGE: clang{{.*}} "-fsycl-is-host"{{.*}} "-o" "[[HOST_OBJECT:.+\.o]]"{{.*}} "[[APPEND_SOURCE]]"
+// CREATE_IMAGE: clang-offload-bundler{{.*}} "-targets=sycl-spir64_gen_image-unknown-unknown,host-x86_64-unknown-linux-gnu" "-output={{.*}}" "-input=[[DEVICE_OBJECT]]" "-input=[[HOST_OBJECT]]"
+
+// RUN: %clangxx -c -fno-sycl-rdc -fsycl -fsycl-targets=spir64_gen \
+// RUN:          --target=x86_64-unknown-linux-gnu -Xsycl-target-backend \
+// RUN:          "-device skl" --sysroot=%S/Inputs/SYCL -ccc-print-phases %s \
+// RUN:          -fno-sycl-device-lib=all 2>&1 \
+// RUN:  | FileCheck %s -check-prefix=CREATE_IMAGE_PHASES
+// CREATE_IMAGE_PHASES: 0: input, "[[INPUT:.+\.cpp]]", c++, (device-sycl)
+// CREATE_IMAGE_PHASES: 1: preprocessor, {0}, c++-cpp-output, (device-sycl)
+// CREATE_IMAGE_PHASES: 2: compiler, {1}, ir, (device-sycl)
+// CREATE_IMAGE_PHASES: 3: input, "{{.*libsycl-itt-user-wrappers.o.*}}", object
+// CREATE_IMAGE_PHASES: 4: clang-offload-unbundler, {3}, object
+// CREATE_IMAGE_PHASES: 5: offload, " (spir64_gen-unknown-unknown)" {4}, object
+// CREATE_IMAGE_PHASES: 6: input, "{{.*libsycl-itt-compiler-wrappers.o.*}}", object
+// CREATE_IMAGE_PHASES: 7: clang-offload-unbundler, {6}, object
+// CREATE_IMAGE_PHASES: 8: offload, " (spir64_gen-unknown-unknown)" {7}, object
+// CREATE_IMAGE_PHASES: 9: input, "{{.*libsycl-itt-stubs.o.*}}", object
+// CREATE_IMAGE_PHASES: 10: clang-offload-unbundler, {9}, object
+// CREATE_IMAGE_PHASES: 11: offload, " (spir64_gen-unknown-unknown)" {10}, object
+// CREATE_IMAGE_PHASES: 12: linker, {5, 8, 11}, ir, (device-sycl)
+// CREATE_IMAGE_PHASES: 13: linker, {2, 12}, ir, (device-sycl)
+// CREATE_IMAGE_PHASES: 14: sycl-post-link, {13}, tempfiletable, (device-sycl)
+// CREATE_IMAGE_PHASES: 15: file-table-tform, {14}, tempfilelist, (device-sycl)
+// CREATE_IMAGE_PHASES: 16: llvm-spirv, {15}, tempfilelist, (device-sycl)
+// CREATE_IMAGE_PHASES: 17: backend-compiler, {16}, image, (device-sycl)
+// CREATE_IMAGE_PHASES: 18: file-table-tform, {14, 17}, tempfiletable, (device-sycl)
+// CREATE_IMAGE_PHASES: 19: clang-offload-wrapper, {18}, object, (device-sycl)
+// CREATE_IMAGE_PHASES: 20: offload, "device-sycl (spir64_gen-unknown-unknown)" {19}, object
+// CREATE_IMAGE_PHASES: 21: offload, "device-sycl (spir64_gen-unknown-unknown)" {20}, object
+// CREATE_IMAGE_PHASES: 22: input, "[[INPUT]]", c++, (host-sycl)
+// CREATE_IMAGE_PHASES: 23: append-footer, {22}, c++, (host-sycl)
+// CREATE_IMAGE_PHASES: 24: preprocessor, {23}, c++-cpp-output, (host-sycl)
+// CREATE_IMAGE_PHASES: 25: offload, "host-sycl (x86_64-unknown-linux-gnu)" {24}, "device-sycl (spir64_gen-unknown-unknown)" {20}, c++-cpp-output
+// CREATE_IMAGE_PHASES: 26: compiler, {25}, ir, (host-sycl)
+// CREATE_IMAGE_PHASES: 27: backend, {26}, assembler, (host-sycl)
+// CREATE_IMAGE_PHASES: 28: assembler, {27}, object, (host-sycl)
+// CREATE_IMAGE_PHASES: 29: clang-offload-bundler, {21, 28}, object, (host-sycl)
+
+// Use of -fno-sycl-rdc -c with non-AOT should not perform the device link.
+// RUN: %clangxx -c -fno-sycl-rdc -fsycl -fsycl-targets=spir64 \
+// RUN:          --target=x86_64-unknown-linux-gnu -ccc-print-phases %s \
+// RUN:          -fno-sycl-device-lib=all 2>&1 \
+// RUN:  | FileCheck %s -check-prefix=JIT_ONLY_PHASES
+// JIT_ONLY_PHASES: 0: input, "[[INPUT:.+\.cpp]]", c++, (device-sycl)
+// JIT_ONLY_PHASES: 1: preprocessor, {0}, c++-cpp-output, (device-sycl)
+// JIT_ONLY_PHASES: 2: compiler, {1}, ir, (device-sycl)
+// JIT_ONLY_PHASES: 3: offload, "device-sycl (spir64-unknown-unknown)" {2}, ir
+// JIT_ONLY_PHASES: 4: input, "[[INPUT]]", c++, (host-sycl)
+// JIT_ONLY_PHASES: 5: append-footer, {4}, c++, (host-sycl)
+// JIT_ONLY_PHASES: 6: preprocessor, {5}, c++-cpp-output, (host-sycl)
+// JIT_ONLY_PHASES: 7: offload, "host-sycl (x86_64-unknown-linux-gnu)" {6}, "device-sycl (spir64-unknown-unknown)" {2}, c++-cpp-output
+// JIT_ONLY_PHASES: 8: compiler, {7}, ir, (host-sycl)
+// JIT_ONLY_PHASES: 9: backend, {8}, assembler, (host-sycl)
+// JIT_ONLY_PHASES: 10: assembler, {9}, object, (host-sycl)
+// JIT_ONLY_PHASES: 11: clang-offload-bundler, {3, 10}, object, (host-sycl)
+
+// Mix and match JIT and AOT phases check.  Expectation is for AOT to perform
+// early device link, and JIT to just produce the LLVM-IR.
+// RUN: %clangxx -c -fno-sycl-rdc -fsycl -fsycl-targets=spir64,spir64_gen \
+// RUN:          --target=x86_64-unknown-linux-gnu --sysroot=%S/Inputs/SYCL \
+// RUN:          -Xsycl-target-backend=spir64_gen "-device skl" \
+// RUN:          -ccc-print-phases %s -fno-sycl-device-lib=all 2>&1 \
+// RUN:  | FileCheck %s -check-prefix=JIT_AOT_PHASES
+// JIT_AOT_PHASES: 0: input, "[[INPUT:.+\.cpp]]", c++, (device-sycl)
+// JIT_AOT_PHASES: 1: preprocessor, {0}, c++-cpp-output, (device-sycl)
+// JIT_AOT_PHASES: 2: compiler, {1}, ir, (device-sycl)
+// JIT_AOT_PHASES: 3: offload, "device-sycl (spir64-unknown-unknown)" {2}, ir
+// JIT_AOT_PHASES: 4: input, "[[INPUT]]", c++, (device-sycl)
+// JIT_AOT_PHASES: 5: preprocessor, {4}, c++-cpp-output, (device-sycl)
+// JIT_AOT_PHASES: 6: compiler, {5}, ir, (device-sycl)
+// JIT_AOT_PHASES: 7: input, "{{.*libsycl-itt-user-wrappers.o.*}}", object
+// JIT_AOT_PHASES: 8: clang-offload-unbundler, {7}, object
+// JIT_AOT_PHASES: 9: offload, " (spir64_gen-unknown-unknown)" {8}, object
+// JIT_AOT_PHASES: 10: input, "{{.*libsycl-itt-compiler-wrappers.o.*}}", object
+// JIT_AOT_PHASES: 11: clang-offload-unbundler, {10}, object
+// JIT_AOT_PHASES: 12: offload, " (spir64_gen-unknown-unknown)" {11}, object
+// JIT_AOT_PHASES: 13: input, "{{.*libsycl-itt-stubs.o.*}}", object
+// JIT_AOT_PHASES: 14: clang-offload-unbundler, {13}, object
+// JIT_AOT_PHASES: 15: offload, " (spir64_gen-unknown-unknown)" {14}, object
+// JIT_AOT_PHASES: 16: linker, {9, 12, 15}, ir, (device-sycl)
+// JIT_AOT_PHASES: 17: linker, {6, 16}, ir, (device-sycl)
+// JIT_AOT_PHASES: 18: sycl-post-link, {17}, tempfiletable, (device-sycl)
+// JIT_AOT_PHASES: 19: file-table-tform, {18}, tempfilelist, (device-sycl)
+// JIT_AOT_PHASES: 20: llvm-spirv, {19}, tempfilelist, (device-sycl)
+// JIT_AOT_PHASES: 21: backend-compiler, {20}, image, (device-sycl)
+// JIT_AOT_PHASES: 22: file-table-tform, {18, 21}, tempfiletable, (device-sycl)
+// JIT_AOT_PHASES: 23: clang-offload-wrapper, {22}, object, (device-sycl)
+// JIT_AOT_PHASES: 24: offload, "device-sycl (spir64_gen-unknown-unknown)" {23}, object
+// JIT_AOT_PHASES: 25: offload, "device-sycl (spir64_gen-unknown-unknown)" {24}, object
+// JIT_AOT_PHASES: 26: input, "[[INPUT]]", c++, (host-sycl)
+// JIT_AOT_PHASES: 27: append-footer, {26}, c++, (host-sycl)
+// JIT_AOT_PHASES: 28: preprocessor, {27}, c++-cpp-output, (host-sycl)
+// JIT_AOT_PHASES: 29: offload, "host-sycl (x86_64-unknown-linux-gnu)" {28}, "device-sycl (spir64_gen-unknown-unknown)" {24}, c++-cpp-output
+// JIT_AOT_PHASES: 30: compiler, {29}, ir, (host-sycl)
+// JIT_AOT_PHASES: 31: backend, {30}, assembler, (host-sycl)
+// JIT_AOT_PHASES: 32: assembler, {31}, object, (host-sycl)
+// JIT_AOT_PHASES: 33: clang-offload-bundler, {3, 25, 32}, object, (host-sycl)
+
+// Consume object and library that contain final device images.
+// RUN: %clangxx -fsycl --target=x86_64-unknown-linux-gnu -### \
+// RUN:          %S/Inputs/SYCL/objgenimage.o %s 2>&1 \
+// RUN:  | FileCheck %s -check-prefix=CONSUME_OBJ
+// CONSUME_OBJ: clang-offload-bundler{{.*}} "-type=o" "-targets=sycl-spir64_gen_image-unknown-unknown" "-input={{.*}}objgenimage.o" "-output=[[DEVICE_IMAGE_OBJ:.+\.o]]
+// CONSUME_OBJ: ld{{.*}} "[[DEVICE_IMAGE_OBJ]]"
+
+// RUN: %clangxx -fsycl --target=x86_64-unknown-linux-gnu -### \
+// RUN:          %S/Inputs/SYCL/libgenimage.a  %s 2>&1 \
+// RUN:  | FileCheck %s -check-prefix=CONSUME_LIB
+// CONSUME_LIB: clang-offload-bundler{{.*}} "-type=aoo" "-targets=sycl-spir64_gen_image-unknown-unknown" "-input={{.*}}libgenimage.a" "-output=[[DEVICE_IMAGE_LIB:.+\.txt]]
+// CONSUME_LIB: ld{{.*}} "@[[DEVICE_IMAGE_LIB]]"

llvm/include/llvm/TargetParser/Triple.h

@@ -156,8 +156,11 @@ class Triple {
     MipsSubArch_r6,
 
     SPIRSubArch_fpga,
+    SPIRSubArch_fpga_image,
     SPIRSubArch_gen,
+    SPIRSubArch_gen_image,
     SPIRSubArch_x86_64,
+    SPIRSubArch_x86_64_image,
 
     PPCSubArch_spe,
 
@@ -785,6 +788,13 @@ class Triple {
     return getArch() == Triple::spir || getArch() == Triple::spir64;
   }
 
+  /// Tests whether the target is SPIR and AOT related.
+  bool isSPIRAOT() const {
+    return isSPIR() && (getSubArch() == Triple::SPIRSubArch_fpga ||
+                        getSubArch() == Triple::SPIRSubArch_gen ||
+                        getSubArch() == Triple::SPIRSubArch_x86_64);
+  }
+
   /// Tests whether the target is SPIR-V (32/64-bit/Logical).
   bool isSPIRV() const {
     return getArch() == Triple::spirv32 || getArch() == Triple::spirv64 ||

llvm/lib/TargetParser/Triple.cpp

@@ -694,10 +694,16 @@ static Triple::SubArchType parseSubArch(StringRef SubArchName) {
     if (SA.consume_front("spir64_") || SA.consume_front("spir_")) {
       if (SA == "fpga")
         return Triple::SPIRSubArch_fpga;
+      else if (SA == "fpga_image")
+        return Triple::SPIRSubArch_fpga_image;
       else if (SA == "gen")
         return Triple::SPIRSubArch_gen;
+      else if (SA == "gen_image")
+        return Triple::SPIRSubArch_gen_image;
       else if (SA == "x86_64")
         return Triple::SPIRSubArch_x86_64;
+      else if (SA == "x86_64_image")
+        return Triple::SPIRSubArch_x86_64_image;
     }
   }
 

sycl/doc/UsersManual.md

@@ -22,6 +22,14 @@ and not recommended to use in production environment.
     You can specify more than one target, comma separated. Default just in time
     (JIT) compilation target can be added to the list to produce a combination
     of AOT and JIT code in the resulting fat binary.
+
+    Normally, '-fsycl-targets' is specified when linking an application, in
+    which case the AOT compiled device binaries are embedded within the
+    application’s fat executable.  However, this option may also be used in
+    combination with '-c' and '-fno-sycl-rdc' when compiling a source file.
+    In this case, the AOT compiled device binaries are embedded within the fat
+    object file.
+
     The following triples are supported by default:
     * spir64 - this is the default generic SPIR-V target;
     * spir64_x86_64 - generate code ahead of time for x86_64 CPUs;

sycl/doc/design/CompilerAndRuntimeDesign.md

@@ -423,6 +423,30 @@ Case 1 can be identified in the device binary generation stage (step 1) by
 scanning the known kernels. Case 2 must be verified by the driver by checking
 for newly introduced kernels in the final link stage (step 3).
 
+#### Device Link during compilation
+
+The `-fno-sycl-rdc` flag can be used in combination with the `-c` option
+when generating fat objects. This option combination informs the compiler to
+perform a full device link stage against the device object, creating a fat
+object that contains the corresponding host object and a fully compiled device
+binary. It is expected that usage of `-fno-sycl-rdc` coincide with
+ahead of time compiling.
+
+When using the generated fat object in this case, the compiler will recognize
+the fat object that contains the fully linked device binary. The device binary
+will be unbundled and linked during the final host link and will not be sent
+through any additional device linking steps.
+
+1. Generation of fat object: a.cpp -> a_fat.o (contains host object and full
+device image)
+2. Linking: a_fat.o -> executable
+
+The generation of the full device image during the compilation (-c) step of
+creating the object allows for library creation that does not require full
+device linking steps which can be a burden to the user.  Providing these early
+device linking steps give the provider of the archives/objects a better user
+experience.
+
 #### Device code post-link step
 
 At link time all the device code is linked into