[DAE][SYCL] Emit MD instead of updating integration header

clang/include/clang/Basic/LangOptions.def

@@ -251,6 +251,7 @@ LANGOPT(SYCLUnnamedLambda , 1, 0, "Allow unnamed lambda SYCL kernels")
 LANGOPT(SYCLVersion       , 32, 0, "Version of the SYCL standard used")
 LANGOPT(DeclareSPIRVBuiltins, 1, 0, "Declare SPIR-V builtin functions")
 LANGOPT(SYCLExplicitSIMD  , 1, 0, "SYCL compilation with explicit SIMD extension")
+LANGOPT(EnableDAEInSpirKernels , 1, 0, "Enable Dead Argument Elimination in SPIR kernels")
 
 LANGOPT(HIPUseNewLaunchAPI, 1, 0, "Use new kernel launching API for HIP")
 

clang/include/clang/Driver/Options.td

@@ -4455,6 +4455,8 @@ def fsycl_std_layout_kernel_params: Flag<["-"], "fsycl-std-layout-kernel-params"
 def fsycl_allow_func_ptr : Flag<["-"], "fsycl-allow-func-ptr">,
   HelpText<"Allow function pointers in SYCL device.">;
 def fno_sycl_allow_func_ptr : Flag<["-"], "fno-sycl-allow-func-ptr">;
+def fenable_sycl_dae : Flag<["-"], "fenable-sycl-dae">,
+  HelpText<"Enable Dead Argument Elimination in SPIR kernels">;
 
 } // let Flags = [CC1Option]
 

clang/lib/CodeGen/BackendUtil.cpp

@@ -918,6 +918,15 @@ void EmitAssemblyHelper::EmitAssembly(BackendAction Action,
   if (LangOpts.SYCLIsDevice && CodeGenOpts.DisableLLVMPasses)
     PerModulePasses.add(createDeadCodeEliminationPass());
 
+  // Eliminate dead arguments from SPIR kernels in SYCL environment.
+  // 1. Run DAE when LLVM optimizations are applied as well.
+  // 2. We cannot run DAE for ESIMD since the pointers to SPIR kernel
+  //    functions are saved in !genx.kernels metadata.
+  // 3. DAE pass temporary guarded under option.
+  if (LangOpts.SYCLIsDevice && !CodeGenOpts.DisableLLVMPasses &&
+      !LangOpts.SYCLExplicitSIMD && LangOpts.EnableDAEInSpirKernels)
+    PerModulePasses.add(createDeadArgEliminationSYCLPass());
+
   if (LangOpts.SYCLIsDevice && LangOpts.SYCLExplicitSIMD)
     PerModulePasses.add(createGenXSPIRVWriterAdaptorPass());
 

clang/lib/Frontend/CompilerInvocation.cpp

@@ -2596,6 +2596,7 @@ static void ParseLangArgs(LangOptions &Opts, ArgList &Args, InputKind IK,
       }
     }
     Opts.SYCLExplicitSIMD = Args.hasArg(options::OPT_fsycl_esimd);
+    Opts.EnableDAEInSpirKernels = Args.hasArg(options::OPT_fenable_sycl_dae);
   }
 
   Opts.IncludeDefaultHeader = Args.hasArg(OPT_finclude_default_header);

llvm/lib/Transforms/IPO/DeadArgumentElimination.cpp

@@ -41,7 +41,6 @@
 #include "llvm/InitializePasses.h"
 #include "llvm/Pass.h"
 #include "llvm/Support/Casting.h"
-#include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/raw_ostream.h"
 #include "llvm/Transforms/IPO.h"
@@ -55,11 +54,6 @@ using namespace llvm;
 
 #define DEBUG_TYPE "deadargelim"
 
-static cl::opt<std::string>
-    IntegrationHeaderFileName("integr-header-file",
-                              cl::desc("Path to integration header file"),
-                              cl::value_desc("filename"), cl::Hidden);
-
 STATISTIC(NumArgumentsEliminated, "Number of unread args removed");
 STATISTIC(NumRetValsEliminated  , "Number of unused return values removed");
 STATISTIC(NumArgumentsReplacedWithUndef,
@@ -760,78 +754,6 @@ void DeadArgumentEliminationPass::PropagateLiveness(const RetOrArg &RA) {
   Uses.erase(Begin, I);
 }
 
-// Update kernel arguments table inside the integration header.
-// For example:
-//   static constexpr const bool param_omit_table[] = {
-//     // OMIT_TABLE_BEGIN
-//     // kernel_name_1
-//     false, false,     // <= update to true if the argument is dead
-//     // kernel_name_2
-//     false, false,
-//     // OMIT_TABLE_END
-//   };
-//   TODO: batch changes to multiple SPIR kernels and do one bulk update.
-constexpr StringLiteral OMIT_TABLE_BEGIN("// OMIT_TABLE_BEGIN");
-constexpr StringLiteral OMIT_TABLE_END("// OMIT_TABLE_END");
-static void updateIntegrationHeader(StringRef SpirKernelName,
-                                    const ArrayRef<bool> &ArgAlive) {
-  ErrorOr<std::unique_ptr<MemoryBuffer>> IntHeaderBuffer =
-      MemoryBuffer::getFile(IntegrationHeaderFileName);
-
-  if (!IntHeaderBuffer)
-    report_fatal_error("unable to read integration header file '" +
-                       IntegrationHeaderFileName +
-                       "': " + IntHeaderBuffer.getError().message());
-
-  // 1. Find the region between OMIT_TABLE_BEGIN and OMIT_TABLE_END
-  StringRef IntHeader((*IntHeaderBuffer)->getBuffer());
-  if (!IntHeader.contains(OMIT_TABLE_BEGIN))
-    report_fatal_error(OMIT_TABLE_BEGIN +
-                       " marker not found in integration header");
-  if (!IntHeader.contains(OMIT_TABLE_END))
-    report_fatal_error(OMIT_TABLE_END +
-                       " marker not found in integration header");
-
-  size_t BeginRegionPos =
-      IntHeader.find(OMIT_TABLE_BEGIN) + OMIT_TABLE_BEGIN.size();
-  size_t EndRegionPos = IntHeader.find(OMIT_TABLE_END);
-
-  StringRef OmitArgTable = IntHeader.slice(BeginRegionPos, EndRegionPos);
-
-  // 2. Find the line that corresponds to the SPIR kernel
-  if (!OmitArgTable.contains(SpirKernelName))
-    report_fatal_error(
-        "Argument table not found in integration header for function '" +
-        SpirKernelName + "'");
-
-  size_t BeginLinePos =
-      OmitArgTable.find(SpirKernelName) + SpirKernelName.size();
-  size_t EndLinePos = OmitArgTable.find("//", BeginLinePos);
-
-  StringRef OmitArgLine = OmitArgTable.slice(BeginLinePos, EndLinePos);
-
-  size_t LineLeftTrim = OmitArgLine.size() - OmitArgLine.ltrim().size();
-  size_t LineRightTrim = OmitArgLine.size() - OmitArgLine.rtrim().size();
-
-  // 3. Construct new file contents and replace only that string.
-  std::string NewIntHeader;
-  NewIntHeader +=
-      IntHeader.take_front(BeginRegionPos + BeginLinePos + LineLeftTrim);
-  for (auto &AliveArg : ArgAlive)
-    NewIntHeader += AliveArg ? "false, " : "true, ";
-  NewIntHeader += IntHeader.drop_front(BeginRegionPos + BeginLinePos +
-                                       OmitArgLine.size() - LineRightTrim);
-
-  // 4. Flush the string into the file.
-  std::error_code EC;
-  raw_fd_ostream File(IntegrationHeaderFileName, EC, sys::fs::F_Text);
-
-  if (EC)
-    report_fatal_error("Cannot open integration header for writing.");
-
-  File << NewIntHeader;
-}
-
 // RemoveDeadStuffFromFunction - Remove any arguments and return values from F
 // that are not in LiveValues. Transform the function and all of the callees of
 // the function to not have these arguments and return values.
@@ -875,8 +797,17 @@ bool DeadArgumentEliminationPass::RemoveDeadStuffFromFunction(Function *F) {
     }
   }
 
-  if (CheckSpirKernels)
-    updateIntegrationHeader(F->getName(), ArgAlive);
+  if (CheckSpirKernels) {
+    SmallVector<Metadata *, 10> MDOmitArgs;
+    auto MDOmitArgTrue = llvm::ConstantAsMetadata::get(
+        ConstantInt::get(Type::getInt1Ty(F->getContext()), 1));
+    auto MDOmitArgFalse = llvm::ConstantAsMetadata::get(
+        ConstantInt::get(Type::getInt1Ty(F->getContext()), 0));
+    for (auto &AliveArg : ArgAlive)
+      MDOmitArgs.push_back(AliveArg ? MDOmitArgFalse : MDOmitArgTrue);
+    F->setMetadata("spir_kernel_omit_args",
+                   llvm::MDNode::get(F->getContext(), MDOmitArgs));
+  }
 
   // Find out the new return value.
   Type *RetTy = FTy->getReturnType();
@@ -1193,11 +1124,6 @@ bool DeadArgumentEliminationPass::RemoveDeadStuffFromFunction(Function *F) {
 
 PreservedAnalyses DeadArgumentEliminationPass::run(Module &M,
                                                    ModuleAnalysisManager &) {
-  // Integration header file must be provided for
-  // DAE to work on SPIR kernels.
-  if (CheckSpirKernels && !IntegrationHeaderFileName.getNumOccurrences())
-    return PreservedAnalyses::all();
-
   bool Changed = false;
 
   // First pass: Do a simple check to see if any functions can have their "..."

llvm/lib/Transforms/IPO/PassManagerBuilder.cpp

@@ -680,7 +680,6 @@ void PassManagerBuilder::populateModulePassManager(
   if (RunInliner) {
     MPM.add(createGlobalOptimizerPass());
     MPM.add(createGlobalDCEPass());
-    MPM.add(createDeadArgEliminationSYCLPass());
   }
 
   // If we are planning to perform ThinLTO later, let's not bloat the code with

llvm/test/Other/opt-O2-pipeline.ll

@@ -193,7 +193,6 @@
 ; CHECK-NEXT:         Branch Probability Analysis
 ; CHECK-NEXT:         Block Frequency Analysis
 ; CHECK-NEXT:     Dead Global Elimination
-; CHECK-NEXT:     Dead Argument Elimination for SPIR kernels in SYCL environment
 ; CHECK-NEXT:     CallGraph Construction
 ; CHECK-NEXT:     Globals Alias Analysis
 ; CHECK-NEXT:     FunctionPass Manager

llvm/test/Other/opt-O3-pipeline-enable-matrix.ll

@@ -198,7 +198,6 @@
 ; CHECK-NEXT:         Branch Probability Analysis
 ; CHECK-NEXT:         Block Frequency Analysis
 ; CHECK-NEXT:     Dead Global Elimination
-; CHECK-NEXT:     Dead Argument Elimination for SPIR kernels in SYCL environment
 ; CHECK-NEXT:     CallGraph Construction
 ; CHECK-NEXT:     Globals Alias Analysis
 ; CHECK-NEXT:     FunctionPass Manager

llvm/test/Other/opt-O3-pipeline.ll

@@ -198,7 +198,6 @@
 ; CHECK-NEXT:         Branch Probability Analysis
 ; CHECK-NEXT:         Block Frequency Analysis
 ; CHECK-NEXT:     Dead Global Elimination
-; CHECK-NEXT:     Dead Argument Elimination for SPIR kernels in SYCL environment
 ; CHECK-NEXT:     CallGraph Construction
 ; CHECK-NEXT:     Globals Alias Analysis
 ; CHECK-NEXT:     FunctionPass Manager

llvm/test/Other/opt-Os-pipeline.ll

@@ -179,7 +179,6 @@
 ; CHECK-NEXT:         Branch Probability Analysis
 ; CHECK-NEXT:         Block Frequency Analysis
 ; CHECK-NEXT:     Dead Global Elimination
-; CHECK-NEXT:     Dead Argument Elimination for SPIR kernels in SYCL environment
 ; CHECK-NEXT:     CallGraph Construction
 ; CHECK-NEXT:     Globals Alias Analysis
 ; CHECK-NEXT:     FunctionPass Manager

llvm/test/Transforms/DeadArgElim/sycl-kernels-neg5.ll → llvm/test/Transforms/DeadArgElim/sycl-kernels-neg.ll

@@ -3,7 +3,9 @@
 ; RUN: opt < %s -deadargelim-sycl -S | FileCheck %s
 
 ; This test ensures dead arguments are not eliminated
-; from a global function that is not a SYCL kernel.
+; from a global function that is not a SPIR kernel.
+
+; CHECK-NOT:    !spir_kernel_omit_args
 
 target triple = "spir64-unknown-unknown-sycldevice"
 

llvm/test/Transforms/DeadArgElim/sycl-kernels.ll

@@ -1,17 +1,6 @@
 ; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --function-signature
-; RUN: echo 'some code we want to preserve' > %t-int_header.h
-; RUN: echo 'static constexpr const bool param_omit_table[] = {' >> %t-int_header.h
-; RUN: echo '  // OMIT_TABLE_BEGIN' >> %t-int_header.h
-; RUN: echo '    // SpirKernel1' >> %t-int_header.h
-; RUN: echo '    false, false,' >> %t-int_header.h
-; RUN: echo '    // SpirKernel2' >> %t-int_header.h
-; RUN: echo '    false, false,' >> %t-int_header.h
-; RUN: echo '  // OMIT_TABLE_END' >> %t-int_header.h
-; RUN: echo '};' >> %t-int_header.h
-; RUN: echo 'some code we want to preserve' >> %t-int_header.h
 ; RUN: opt < %s -deadargelim -S | FileCheck %s --check-prefixes=CHECK
-; RUN: opt < %s -deadargelim-sycl -S -integr-header-file %t-int_header.h | FileCheck %s --check-prefixes=CHECK-SYCL
-; RUN: cat %t-int_header.h | FileCheck %s --check-prefixes=CHECK-INT-HEADER
+; RUN: opt < %s -deadargelim-sycl -S | FileCheck %s --check-prefixes=CHECK-SYCL
 
 ; This test checks eliminating dead arguments
 ; from SPIR kernel functions in SYCL environment.
@@ -25,10 +14,10 @@ define weak_odr spir_kernel void @SpirKernel1(float %arg1, float %arg2) {
 ; CHECK-NEXT:    ret void
 ;
 ; CHECK-SYCL-LABEL: define {{[^@]+}}@SpirKernel1
-; CHECK-SYCL-SAME: (float [[ARG1:%.*]])
+; CHECK-SYCL-SAME: (float [[ARG1:%.*]]) !spir_kernel_omit_args ![[KERN_ARGS1:[0-9]]]
 ; CHECK-SYCL-NEXT:    call void @foo(float [[ARG1]])
 ; CHECK-SYCL-NEXT:    ret void
-;
+
   call void @foo(float %arg1)
   ret void
 }
@@ -40,23 +29,15 @@ define weak_odr spir_kernel void @SpirKernel2(float %arg1, float %arg2) {
 ; CHECK-NEXT:    ret void
 ;
 ; CHECK-SYCL-LABEL: define {{[^@]+}}@SpirKernel2
-; CHECK-SYCL-SAME: (float [[ARG2:%.*]])
+; CHECK-SYCL-SAME: (float [[ARG2:%.*]]) !spir_kernel_omit_args ![[KERN_ARGS2:[0-9]]]
 ; CHECK-SYCL-NEXT:    call void @foo(float [[ARG2]])
 ; CHECK-SYCL-NEXT:    ret void
-;
+
   call void @foo(float %arg2)
   ret void
 }
 
-; CHECK-INT-HEADER: some code we want to preserve
-; CHECK-INT-HEADER: static constexpr const bool param_omit_table[] = {
-; CHECK-INT-HEADER:   // OMIT_TABLE_BEGIN
-; CHECK-INT-HEADER:     // SpirKernel1
-; CHECK-INT-HEADER:     false, true,
-; CHECK-INT-HEADER:     // SpirKernel2
-; CHECK-INT-HEADER:     true, false,
-; CHECK-INT-HEADER:   // OMIT_TABLE_END
-; CHECK-INT-HEADER: };
-; CHECK-INT-HEADER: some code we want to preserve
+; CHECK-SYCL: ![[KERN_ARGS1]] = !{i1 false, i1 true}
+; CHECK-SYCL: ![[KERN_ARGS2]] = !{i1 true, i1 false}
 
 declare void @foo(float %arg)