[SYCL] Propagate SPIRV kernel parameter opt info to runtime.

The post-link tool
- uses a new analysis pass (SPIRKernelParamOptInfo) to extract kernel
  parameter optimization info from IR produced by the DeadArgumentElimination
  pass
- uses the PropertySetIO mechanism to encode this information into the ourput
  property file
Then the clang offload wrapper picks up the property file and injects it into
the device binary descriptor accessed by the SYCL runtime when the app runs.

Signed-off-by: Konstantin S Bobrovsky <[email protected]>

Author: kbobrovs

llvm/include/llvm/Support/PropertySetIO.h

@@ -180,6 +180,7 @@ class PropertySetRegistry {
   static constexpr char SYCL_SPECIALIZATION_CONSTANTS[] =
       "SYCL/specialization constants";
   static constexpr char SYCL_DEVICELIB_REQ_MASK[] = "SYCL/devicelib req mask";
+  static constexpr char SYCL_KERNEL_PARAM_OPT_INFO[] = "SYCL/kernel param opt";
 
   // Function for bulk addition of an entire property set under given category
   // (property set name).

llvm/test/tools/sycl-post-link/omit_kernel_args.ll

@@ -0,0 +1,44 @@
+; This test checks that the post-link tool generates correct kernel parameter
+; optimization info into a property file if the source IR contained
+; corresponding metadata.
+;
+; RUN: sycl-post-link -emit-param-info -S %s -o %t.files.table
+; RUN: FileCheck %s -input-file=%t.files.table --check-prefixes CHECK-TABLE
+; RUN: FileCheck %s -input-file=%t.files_0.prop --match-full-lines --check-prefixes CHECK-PROP
+
+target triple = "spir64-unknown-unknown-sycldevice"
+
+define weak_odr spir_kernel void @SpirKernel1(float %arg1) !spir_kernel_omit_args !0 {
+  call void @foo(float %arg1)
+  ret void
+}
+
+define weak_odr spir_kernel void @SpirKernel2(i8 %arg1, i8 %arg2, i8 %arg3) !spir_kernel_omit_args !1 {
+  call void @bar(i8 %arg1)
+  call void @bar(i8 %arg2)
+  call void @bar(i8 %arg3)
+  ret void
+}
+
+declare void @foo(float)
+declare void @bar(i8)
+
+; // HEX: 0x01 (1 byte)
+; // BIN: 01 - 2 arguments total, 1 remains
+; // NOTE: In the metadata string below, unlike bitmask above, least significant
+; //       element goes first
+!0 = !{i1 true, i1 false}
+
+; // HEX: 0x7E 0x06 (2 bytes)
+; // BIN: 110 01111110 - 11 arguments total, 3 remain
+!1 = !{i1 false, i1 true, i1 true, i1 true, i1 true, i1 true, i1 true, i1 false, i1 false, i1 true, i1 true}
+
+; CHECK-PROP: [SYCL/kernel param opt]
+; // Base64 encoding in the prop file (including 8 bytes length):
+; CHECK-PROP-NEXT: SpirKernel1=2|CAAAAAAAAAQA
+; // Base64 encoding in the prop file (including 8 bytes length):
+; CHECK-PROP-NEXT: SpirKernel2=2|LAAAAAAAAAgfGA
+
+; CHECK-TABLE: [Code|Properties]
+; CHECK-TABLE-NEXT: {{.*}}files_0.prop
+; CHECK-TABLE-EMPTY:

llvm/tools/sycl-post-link/CMakeLists.txt

@@ -9,6 +9,7 @@ set(LLVM_LINK_COMPONENTS
 
 add_llvm_tool(sycl-post-link
   sycl-post-link.cpp
+  SPIRKernelParamOptInfo.cpp
   SpecConstants.cpp
 
   DEPENDS

llvm/tools/sycl-post-link/SPIRKernelParamOptInfo.cpp

@@ -0,0 +1,51 @@
+//==-- SPIRKernelParamOptInfo.cpp -- get kernel param optimization info ----==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#include "SPIRKernelParamOptInfo.h"
+
+#include "llvm/IR/Constants.h"
+#include "llvm/Support/Casting.h"
+
+namespace {
+
+// Must match the one produced by DeadArgumentElimination
+static constexpr char MetaDataID[] = "spir_kernel_omit_args";
+
+} // anonymous namespace
+
+namespace llvm {
+
+void SPIRKernelParamOptInfo::releaseMemory() { clear(); }
+
+SPIRKernelParamOptInfo
+SPIRKernelParamOptInfoAnalysis::run(Module &M, ModuleAnalysisManager &AM) {
+  SPIRKernelParamOptInfo Res;
+
+  for (const Function &F : M) {
+    MDNode *MD = F.getMetadata(MetaDataID);
+    if (!MD)
+      continue;
+    using BaseTy = SPIRKernelParamOptInfoBaseTy;
+    auto Ins =
+        Res.insert(BaseTy::value_type{F.getName(), BaseTy::mapped_type{}});
+    assert(Ins.second && "duplicate kernel?");
+    BitVector &ParamDropped = Ins.first->second;
+
+    for (unsigned int I = 0; I < MD->getNumOperands(); ++I) {
+      const auto *MDInt1 = cast<ConstantAsMetadata>(MD->getOperand(I));
+      unsigned ID = static_cast<unsigned>(
+          cast<ConstantInt>(MDInt1->getValue())->getValue().getZExtValue());
+      ParamDropped.push_back(ID != 0);
+    }
+  }
+  return Res;
+}
+
+AnalysisKey SPIRKernelParamOptInfoAnalysis::Key;
+
+} // namespace llvm

llvm/tools/sycl-post-link/SPIRKernelParamOptInfo.h

@@ -0,0 +1,47 @@
+//==-- SPIRKernelParamOptInfo.h -- get kernel param optimization info ------==//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+// Analysis pass which collects parameter optimization info for SPIR kernels.
+// Currently the info is whether i'th kernel parameter as emitted by the FE has
+// been optimized away by the LLVM optimizer. The information is produced by
+// DeadArgumentElimination transformation and stored into a specific metadata
+// attached to kernel functions in a module.
+//===----------------------------------------------------------------------===//
+
+#include "llvm/ADT/BitVector.h"
+#include "llvm/ADT/StringRef.h"
+#include "llvm/IR/Module.h"
+#include "llvm/IR/PassManager.h"
+
+#include <map>
+#include <string>
+
+namespace llvm {
+
+// the StringRef key refers to a function name
+using SPIRKernelParamOptInfoBaseTy = std::map<StringRef, BitVector>;
+
+class SPIRKernelParamOptInfo : public SPIRKernelParamOptInfoBaseTy {
+public:
+  void releaseMemory();
+};
+
+class SPIRKernelParamOptInfoAnalysis
+    : public AnalysisInfoMixin<SPIRKernelParamOptInfoAnalysis> {
+  friend AnalysisInfoMixin<SPIRKernelParamOptInfoAnalysis>;
+
+  static AnalysisKey Key;
+
+public:
+  /// Provide the result type for this analysis pass.
+  using Result = SPIRKernelParamOptInfo;
+
+  /// Run the analysis pass over a function and produce BPI.
+  SPIRKernelParamOptInfo run(Module &M, ModuleAnalysisManager &AM);
+};
+
+} // namespace llvm

llvm/tools/sycl-post-link/sycl-post-link.cpp

@@ -13,7 +13,9 @@
 // - specialization constant intrinsic transformation
 //===----------------------------------------------------------------------===//
 
+#include "SPIRKernelParamOptInfo.h"
 #include "SpecConstants.h"
+
 #include "llvm/ADT/SetVector.h"
 #include "llvm/ADT/Triple.h"
 #include "llvm/Bitcode/BitcodeWriterPass.h"
@@ -34,6 +36,7 @@
 #include "llvm/Transforms/IPO.h"
 #include "llvm/Transforms/IPO/GlobalDCE.h"
 #include "llvm/Transforms/Utils/Cloning.h"
+
 #include <memory>
 
 using namespace llvm;
@@ -117,11 +120,16 @@ static cl::opt<SpecConstMode> SpecConstLower{
                    "set spec constants to C++ defaults")),
     cl::cat(PostLinkCat)};
 
+static cl::opt<bool> EmitKernelParamInfo{
+    "emit-param-info", cl::desc("emit kernel parameter optimization info"),
+    cl::cat(PostLinkCat)};
+
 struct ImagePropSaveInfo {
   bool NeedDeviceLibReqMask;
   bool DoSpecConst;
   bool SetSpecConstAtRT;
   bool SpecConstsMet;
+  bool EmitKernelParamInfo;
 };
 // Please update DeviceLibFuncMap if any item is added to or removed from
 // fallback device libraries in libdevice.
@@ -508,6 +516,30 @@ static string_vector saveDeviceImageProperty(
           llvm::util::PropertySetRegistry::SYCL_SPECIALIZATION_CONSTANTS,
           TmpSpecIDMap);
     }
+    if (ImgPSInfo.EmitKernelParamInfo) {
+      // extract kernel parameter optimization info per module
+      ModuleAnalysisManager MAM;
+      // Register required analysis
+      MAM.registerPass([&] { return PassInstrumentationAnalysis(); });
+      // Register the payload analysis
+      MAM.registerPass([&] { return SPIRKernelParamOptInfoAnalysis(); });
+      SPIRKernelParamOptInfo PInfo =
+          MAM.getResult<SPIRKernelParamOptInfoAnalysis>(*ResultModules[I]);
+
+      // convert analysis results into properties and record them
+      llvm::util::PropertySet &Props =
+          PropSet[llvm::util::PropertySetRegistry::SYCL_KERNEL_PARAM_OPT_INFO];
+
+      for (const auto &NameInfoPair : PInfo) {
+        const llvm::BitVector &Bits = NameInfoPair.second;
+        const llvm::ArrayRef<uintptr_t> Arr = NameInfoPair.second.getData();
+        const unsigned char *Data =
+            reinterpret_cast<const unsigned char *>(Arr.begin());
+        llvm::util::PropertyValue::SizeTy DataBitSize = Bits.size();
+        Props.insert(std::make_pair(
+            NameInfoPair.first, llvm::util::PropertyValue(Data, DataBitSize)));
+      }
+    }
     std::error_code EC;
     std::string SCFile = makeResultFileName(".prop", I);
     raw_fd_ostream SCOut(SCFile, EC);
@@ -580,8 +612,9 @@ int main(int argc, char **argv) {
 
   bool DoSplit = SplitMode.getNumOccurrences() > 0;
   bool DoSpecConst = SpecConstLower.getNumOccurrences() > 0;
+  bool DoParamInfo = EmitKernelParamInfo.getNumOccurrences() > 0;
 
-  if (!DoSplit && !DoSpecConst && !DoSymGen) {
+  if (!DoSplit && !DoSpecConst && !DoSymGen && !DoParamInfo) {
     errs() << "no actions specified; try --help for usage info\n";
     return 1;
   }
@@ -595,6 +628,11 @@ int main(int argc, char **argv) {
            << IROutputOnly.ArgStr << "\n";
     return 1;
   }
+  if (IROutputOnly && DoParamInfo) {
+    errs() << "error: -" << EmitKernelParamInfo.ArgStr << " can't be used with"
+           << " -" << IROutputOnly.ArgStr << "\n";
+    return 1;
+  }
   SMDiagnostic Err;
   std::unique_ptr<Module> M = parseIRFile(InputFilename, Err, Context);
   // It is OK to use raw pointer here as we control that it does not outlive M
@@ -664,7 +702,7 @@ int main(int argc, char **argv) {
 
   {
     ImagePropSaveInfo ImgPSInfo = {true, DoSpecConst, SetSpecConstAtRT,
-                                   SpecConstsMet};
+                                   SpecConstsMet, EmitKernelParamInfo};
     string_vector Files = saveDeviceImageProperty(ResultModules, ImgPSInfo);
     Error Err = Table.addColumn(COL_PROPS, Files);
     CHECK_AND_EXIT(Err);