[SYCL][Driver] Expose llvm-foreach --jobs functionality through a driver option (#4543)

AlexeySachkov · Artem Gindinson · web-flow · commit c2221f035d80 · 2021-09-22T11:47:02.000+03:00
After #4360 extended `llvm-foreach` with an option to specify the desired number of parallel jobs, expose this functionality to the DPC++ compiler users. Co-authored-by: Alexey Sachkov <alexey.sachkov@intel.com> Co-authored-by: Artem Gindinson <artem.gindinson@intel.com>
diff --git a/clang/include/clang/Driver/Options.td b/clang/include/clang/Driver/Options.td
@@ -2685,6 +2685,11 @@ def fsycl_footer_path_EQ : Joined<["-"], "fsycl-footer-path=">,
 def fno_sycl_link_spirv : Flag<["-"], "fno-sycl-link-spirv">,
   Flags<[CoreOption]>, HelpText<"Disable adding of the default (spir64) triple "
   "when discovered in user specified objects and archives.">;
+def fsycl_max_parallel_jobs_EQ : Joined<["-"], "fsycl-max-parallel-link-jobs=">,
+  Flags<[CoreOption]>, Group<f_Group>,
+  HelpText<"Experimental feature: Controls the maximum parallelism of actions performed "
+  "on SYCL device code post-link, i.e. the generation of SPIR-V device images "
+  "or AOT compilation of each device image.">;
 def fsyntax_only : Flag<["-"], "fsyntax-only">,
   Flags<[NoXarchOption,CoreOption,CC1Option,FC1Option]>, Group<Action_Group>;
 def ftabstop_EQ : Joined<["-"], "ftabstop=">, Group<f_Group>;
diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp
@@ -8794,6 +8794,11 @@ void SPIRVTranslator::ConstructJob(Compilation &C, const JobAction &JA,
         TCArgs.MakeArgString("--out-file-list=" + OutputFileName));
     ForeachArgs.push_back(
         TCArgs.MakeArgString("--out-replace=" + OutputFileName));
+    StringRef ParallelJobs =
+        TCArgs.getLastArgValue(options::OPT_fsycl_max_parallel_jobs_EQ);
+    if (!ParallelJobs.empty())
+      ForeachArgs.push_back(TCArgs.MakeArgString("--jobs=" + ParallelJobs));
+
     ForeachArgs.push_back(TCArgs.MakeArgString("--"));
     ForeachArgs.push_back(TCArgs.MakeArgString(Cmd->getExecutable()));
 
diff --git a/clang/lib/Driver/ToolChains/SYCL.cpp b/clang/lib/Driver/ToolChains/SYCL.cpp
@@ -97,8 +97,8 @@ void SYCL::constructLLVMForeachCommand(Compilation &C, const JobAction &JA,
                                        std::unique_ptr<Command> InputCommand,
                                        const InputInfoList &InputFiles,
                                        const InputInfo &Output, const Tool *T,
-                                       StringRef Increment,
-                                       StringRef Ext = "out") {
+                                       StringRef Increment, StringRef Ext,
+                                       StringRef ParallelJobs) {
   // Construct llvm-foreach command.
   // The llvm-foreach command looks like this:
   // llvm-foreach --in-file-list=a.list --in-replace='{}' -- echo '{}'
@@ -120,6 +120,9 @@ void SYCL::constructLLVMForeachCommand(Compilation &C, const JobAction &JA,
   if (!Increment.empty())
     ForeachArgs.push_back(
         C.getArgs().MakeArgString("--out-increment=" + Increment));
+  if (!ParallelJobs.empty())
+    ForeachArgs.push_back(C.getArgs().MakeArgString("--jobs=" + ParallelJobs));
+
   ForeachArgs.push_back(C.getArgs().MakeArgString("--"));
   ForeachArgs.push_back(
       C.getArgs().MakeArgString(InputCommand->getExecutable()));
@@ -392,10 +395,12 @@ void SYCL::fpga::BackendCompiler::constructOpenCLAOTCommand(
   const char *Exec = C.getArgs().MakeArgString(ExecPath);
   auto Cmd = std::make_unique<Command>(JA, *this, ResponseFileSupport::None(),
                                        Exec, CmdArgs, None);
-  if (!ForeachInputs.empty())
+  if (!ForeachInputs.empty()) {
+    StringRef ParallelJobs =
+        Args.getLastArgValue(options::OPT_fsycl_max_parallel_jobs_EQ);
     constructLLVMForeachCommand(C, JA, std::move(Cmd), ForeachInputs, Output,
-                                this, "", ForeachExt);
-  else
+                                this, "", ForeachExt, ParallelJobs);
+  } else
     C.addCommand(std::move(Cmd));
 }
 
@@ -557,10 +562,12 @@ void SYCL::fpga::BackendCompiler::ConstructJob(
   auto Cmd = std::make_unique<Command>(JA, *this, ResponseFileSupport::None(),
                                        Exec, CmdArgs, None);
   addFPGATimingDiagnostic(Cmd, C);
-  if (!ForeachInputs.empty())
+  if (!ForeachInputs.empty()) {
+    StringRef ParallelJobs =
+        Args.getLastArgValue(options::OPT_fsycl_max_parallel_jobs_EQ);
     constructLLVMForeachCommand(C, JA, std::move(Cmd), ForeachInputs, Output,
-                                this, ReportOptArg, ForeachExt);
-  else
+                                this, ReportOptArg, ForeachExt, ParallelJobs);
+  } else
     C.addCommand(std::move(Cmd));
 }
 
@@ -596,10 +603,12 @@ void SYCL::gen::BackendCompiler::ConstructJob(Compilation &C,
   const char *Exec = C.getArgs().MakeArgString(ExecPath);
   auto Cmd = std::make_unique<Command>(JA, *this, ResponseFileSupport::None(),
                                        Exec, CmdArgs, None);
-  if (!ForeachInputs.empty())
+  if (!ForeachInputs.empty()) {
+    StringRef ParallelJobs =
+        Args.getLastArgValue(options::OPT_fsycl_max_parallel_jobs_EQ);
     constructLLVMForeachCommand(C, JA, std::move(Cmd), ForeachInputs, Output,
-                                this, "");
-  else
+                                this, "", "out", ParallelJobs);
+  } else
     C.addCommand(std::move(Cmd));
 }
 
@@ -629,10 +638,12 @@ void SYCL::x86_64::BackendCompiler::ConstructJob(
   const char *Exec = C.getArgs().MakeArgString(ExecPath);
   auto Cmd = std::make_unique<Command>(JA, *this, ResponseFileSupport::None(),
                                        Exec, CmdArgs, None);
-  if (!ForeachInputs.empty())
+  if (!ForeachInputs.empty()) {
+    StringRef ParallelJobs =
+        Args.getLastArgValue(options::OPT_fsycl_max_parallel_jobs_EQ);
     constructLLVMForeachCommand(C, JA, std::move(Cmd), ForeachInputs, Output,
-                                this, "");
-  else
+                                this, "", "out", ParallelJobs);
+  } else
     C.addCommand(std::move(Cmd));
 }
 
diff --git a/clang/lib/Driver/ToolChains/SYCL.h b/clang/lib/Driver/ToolChains/SYCL.h
@@ -36,7 +36,8 @@ void constructLLVMForeachCommand(Compilation &C, const JobAction &JA,
                                  std::unique_ptr<Command> InputCommand,
                                  const InputInfoList &InputFiles,
                                  const InputInfo &Output, const Tool *T,
-                                 StringRef Increment, StringRef Ext);
+                                 StringRef Increment, StringRef Ext = "out",
+                                 StringRef ParallelJobs = "");
 
 // Runs llvm-spirv to convert spirv to bc, llvm-link, which links multiple LLVM
 // bitcode. Converts generated bc back to spirv using llvm-spirv, wraps with
diff --git a/clang/test/Driver/sycl-offload-with-split.c b/clang/test/Driver/sycl-offload-with-split.c
@@ -220,6 +220,22 @@
 
 /// ###########################################################################
 
+/// Check parallel compilation enforcement for split modules when running SPIR-V translation and AOT compilation
+// RUN: %clang -target x86_64-unknown-linux-gnu -fsycl -fsycl-max-parallel-link-jobs=4 -fsycl-targets=spir64-unknown-unknown %s -### 2>&1 \
+// RUN:  | FileCheck %s -check-prefixes=CHK-PARALLEL-JOBS
+// RUN: %clang -target x86_64-unknown-linux-gnu -fsycl -fsycl-max-parallel-link-jobs=4 -fsycl-targets=spir64_fpga-unknown-unknown -Xshardware %s -### 2>&1 \
+// RUN:  | FileCheck %s -check-prefixes=CHK-PARALLEL-JOBS,CHK-PARALLEL-JOBS-AOT -DBE_COMPILER=aoc
+// RUN: %clang -target x86_64-unknown-linux-gnu -fsycl -fsycl-max-parallel-link-jobs=4 -fintelfpga -Xshardware %s -### 2>&1 \
+// RUN:  | FileCheck %s -check-prefixes=CHK-PARALLEL-JOBS,CHK-PARALLEL-JOBS-AOT -DBE_COMPILER=aoc
+// RUN: %clang -target x86_64-unknown-linux-gnu -fsycl -fsycl-max-parallel-link-jobs=4 -fsycl-targets=spir64_gen-unknown-unknown %s -### 2>&1 \
+// RUN:  | FileCheck %s -check-prefixes=CHK-PARALLEL-JOBS,CHK-PARALLEL-JOBS-AOT -DBE_COMPILER=ocloc
+// RUN: %clang -target x86_64-unknown-linux-gnu -fsycl -fsycl-max-parallel-link-jobs=4 -fsycl-targets=spir64_x86_64-unknown-unknown %s -### 2>&1 \
+// RUN:  | FileCheck %s -check-prefixes=CHK-PARALLEL-JOBS,CHK-PARALLEL-JOBS-AOT -DBE_COMPILER=opencl-aot
+// CHK-PARALLEL-JOBS: llvm-foreach{{.*}} "--jobs=4" "--" "{{.*}}llvm-spirv{{.*}}"
+// CHK-PARALLEL-JOBS-AOT: llvm-foreach{{.*}} "--jobs=4" "--" "{{.*}}[[BE_COMPILER]]{{.*}}
+
+/// ###########################################################################
+
 /// offload with multiple targets, including AOT
 // RUN:  %clang -target x86_64-unknown-linux-gnu -fsycl -fno-sycl-device-lib=all -fsycl-device-code-split -fsycl-targets=spir64-unknown-unknown,spir64_fpga-unknown-unknown,spir64_gen-unknown-unknown -ccc-print-phases %s 2>&1 \
 // RUN:   | FileCheck -check-prefix=CHK-PHASE-MULTI-TARG %s
diff --git a/sycl/doc/UsersManual.md b/sycl/doc/UsersManual.md
@@ -176,6 +176,15 @@ and not recommended to use in production environment.
     * auto - the compiler will use a heuristic to select the best way of
       splitting device code. This is default mode.
 
+**`-fsycl-max-parallel-link-jobs=<N>`**
+
+    Experimental feature. When specified, it informs the compiler
+    that it can simultaneously spawn up to `N` processes to perform
+    actions required to link the DPC++ application. This option is
+    only useful in SYCL mode. It only takes effect if link action
+    needs to be executed, i.e. it won't have any effect in presence of
+    options like `-c` or `-E`. Default value of `N` is 1.
+
 **`-f[no-]sycl-device-lib=<lib1>[,<lib2>,...]`**
 
     Enables/disables linking of the device libraries. Supported libraries:
