[SYCL] Remove devicelib jit link (#20777)

The '-fsycl-device-lib-jit-link' option has been removed in
#20326. This PR cleans up
sycl-post-link and sycl runtime code to support this option. The removed
code includes:
1. Analysis in sycl-post-link tool to record which device libraries are
used by user's device image via image property
2. Program manager handling for device library require mask property to
decide which fallback spv will be loaded and linked during execution
time.
3. Program manager utils to read/load device library fallback spv files
into memory and utils to link in-memory spv device library modules with
user's module.

---------

Signed-off-by: jinge90 <[email protected]>

Author: jinge90

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

@@ -44,21 +44,18 @@ int main() {
 // CHECK-DAG: @.sycl_offloading.target.0 = internal unnamed_addr constant [7 x i8] c"spir64\00"
 // CHECK-DAG: @.sycl_offloading.opts.compile.0 = internal unnamed_addr constant [1 x i8] zeroinitializer
 // CHECK-DAG: @.sycl_offloading.opts.link.0 = internal unnamed_addr constant [1 x i8] zeroinitializer
-// CHECK-DAG: @prop = internal unnamed_addr constant [17 x i8] c"DeviceLibReqMask\00"
-// CHECK-DAG: @__sycl_offload_prop_sets_arr = internal constant [1 x %_pi_device_binary_property_struct] [%_pi_device_binary_property_struct { ptr @prop, ptr null, i32 1, i64 0 }]
-// CHECK-DAG: @SYCL_PropSetName = internal unnamed_addr constant [24 x i8] c"SYCL/devicelib req mask\00"
-// CHECK-DAG: @prop.1 = internal unnamed_addr constant [8 x i8] c"aspects\00"
+// CHECK-DAG: @prop = internal unnamed_addr constant [8 x i8] c"aspects\00"
 // CHECK-DAG: @prop_val = internal unnamed_addr constant [8 x i8] zeroinitializer
-// CHECK-DAG: @__sycl_offload_prop_sets_arr.2 = internal constant [1 x %_pi_device_binary_property_struct] [%_pi_device_binary_property_struct { ptr @prop.1, ptr @prop_val, i32 2, i64 8 }]
-// CHECK-DAG: @SYCL_PropSetName.3 = internal unnamed_addr constant [25 x i8] c"SYCL/device requirements\00"
-// CHECK-DAG: @SYCL_PropSetName.4 = internal unnamed_addr constant [22 x i8] c"SYCL/kernel param opt\00"
-// CHECK-DAG: @__sycl_offload_prop_sets_arr.5 = internal constant [3 x %_pi_device_binary_property_set_struct] [%_pi_device_binary_property_set_struct { ptr @SYCL_PropSetName, ptr @__sycl_offload_prop_sets_arr, ptr getelementptr ([1 x %_pi_device_binary_property_struct], ptr @__sycl_offload_prop_sets_arr, i64 0, i64 1) }, %_pi_device_binary_property_set_struct { ptr @SYCL_PropSetName.3, ptr @__sycl_offload_prop_sets_arr.2, ptr getelementptr ([1 x %_pi_device_binary_property_struct], ptr @__sycl_offload_prop_sets_arr.2, i64 0, i64 1) }, %_pi_device_binary_property_set_struct { ptr @SYCL_PropSetName.4, ptr null, ptr null }]
+// CHECK-DAG: @__sycl_offload_prop_sets_arr = internal constant [1 x %_pi_device_binary_property_struct] [%_pi_device_binary_property_struct { ptr @prop, ptr @prop_val, i32 2, i64 8 }]
+// CHECK-DAG: @SYCL_PropSetName = internal unnamed_addr constant [25 x i8] c"SYCL/device requirements\00"
+// CHECK-DAG: @SYCL_PropSetName.1 = internal unnamed_addr constant [22 x i8] c"SYCL/kernel param opt\00"
+// CHECK-DAG: @__sycl_offload_prop_sets_arr.2 = internal constant [2 x %_pi_device_binary_property_set_struct] [%_pi_device_binary_property_set_struct { ptr @SYCL_PropSetName, ptr @__sycl_offload_prop_sets_arr, ptr getelementptr ([1 x %_pi_device_binary_property_struct], ptr @__sycl_offload_prop_sets_arr, i64 0, i64 1) }, %_pi_device_binary_property_set_struct { ptr @SYCL_PropSetName.1, ptr null, ptr null }]
 // CHECK-DAG: @.sycl_offloading.0.data = internal unnamed_addr constant [912 x i8]
 // CHECK-DAG: @__sycl_offload_entry_name = internal unnamed_addr constant [25 x i8] c"_ZTSZ4mainE11fake_kernel\00"
 // CHECK-DAG: @__sycl_offload_entries_arr = internal constant [1 x %struct.__tgt_offload_entry] [%struct.__tgt_offload_entry { i64 0, i16 1, i16 8, i32 0, ptr null, ptr @__sycl_offload_entry_name, i64 0, i64 0, ptr null }]
 // CHECK-DAG: @.sycl_offloading.0.info = internal local_unnamed_addr constant [2 x i64] [i64 ptrtoint (ptr @.sycl_offloading.0.data to i64), i64 912], section ".tgtimg", align 16
 // CHECK-DAG: @llvm.used = appending global [1 x ptr] [ptr @.sycl_offloading.0.info], section "llvm.metadata"
-// CHECK-DAG: @.sycl_offloading.device_images = internal unnamed_addr constant [1 x %__sycl.tgt_device_image] [%__sycl.tgt_device_image { i16 3, i8 4, i8 0, ptr @.sycl_offloading.target.0, ptr @.sycl_offloading.opts.compile.0, ptr @.sycl_offloading.opts.link.0, ptr @.sycl_offloading.0.data, ptr getelementptr ([912 x i8], ptr @.sycl_offloading.0.data, i64 0, i64 912), ptr @__sycl_offload_entries_arr, ptr getelementptr ([1 x %struct.__tgt_offload_entry], ptr @__sycl_offload_entries_arr, i64 0, i64 1), ptr @__sycl_offload_prop_sets_arr.5, ptr getelementptr ([3 x %_pi_device_binary_property_set_struct], ptr @__sycl_offload_prop_sets_arr.5, i64 0, i64 3) }]
+// CHECK-DAG: @.sycl_offloading.device_images = internal unnamed_addr constant [1 x %__sycl.tgt_device_image] [%__sycl.tgt_device_image { i16 3, i8 4, i8 0, ptr @.sycl_offloading.target.0, ptr @.sycl_offloading.opts.compile.0, ptr @.sycl_offloading.opts.link.0, ptr @.sycl_offloading.0.data, ptr getelementptr ([912 x i8], ptr @.sycl_offloading.0.data, i64 0, i64 912), ptr @__sycl_offload_entries_arr, ptr getelementptr ([1 x %struct.__tgt_offload_entry], ptr @__sycl_offload_entries_arr, i64 0, i64 1), ptr @__sycl_offload_prop_sets_arr.2, ptr getelementptr ([2 x %_pi_device_binary_property_set_struct], ptr @__sycl_offload_prop_sets_arr.2, i64 0, i64 2) }]
 // CHECK-DAG: @.sycl_offloading.descriptor = internal constant %__sycl.tgt_bin_desc { i16 1, i16 1, ptr @.sycl_offloading.device_images, ptr null, ptr null }
 // CHECK-DAG: @llvm.global_ctors = {{.*}} { i32 1, ptr @sycl.descriptor_reg, ptr null }]
 // CHECK-DAG: @llvm.global_dtors = {{.*}} { i32 1, ptr @sycl.descriptor_unreg, ptr null }]

llvm/include/llvm/SYCLLowerIR/SYCLDeviceLibBF16.h

@@ -0,0 +1,16 @@
+//===----- SYCLDeviceLibBF16.h - get SYCL devicelib required 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
+//
+//===----------------------------------------------------------------------===//
+
+#pragma once
+
+namespace llvm {
+class Function;
+class Module;
+bool isSYCLDeviceLibBF16Used(const Module &M);
+bool isBF16DeviceLibFuncDecl(const Function &F);
+} // namespace llvm

llvm/include/llvm/SYCLLowerIR/SYCLDeviceLibReqMask.h

@@ -11,7 +11,6 @@
 #pragma once
 
 #include "llvm/ADT/SetVector.h"
-#include "llvm/SYCLLowerIR/SYCLDeviceLibReqMask.h"
 #include "llvm/Support/PropertySetIO.h"
 #include <string>
 namespace llvm {

llvm/include/llvm/SYCLPostLink/ComputeModuleRuntimeInfo.h

@@ -209,9 +209,6 @@ class PropertySetRegistry {
       "SYCL/specialization constants";
   static constexpr char SYCL_SPEC_CONSTANTS_DEFAULT_VALUES[] =
       "SYCL/specialization constants default values";
-  // TODO: remove SYCL_DEVICELIB_REQ_MASK when devicelib online linking path
-  // is totally removed.
-  static constexpr char SYCL_DEVICELIB_REQ_MASK[] = "SYCL/devicelib req mask";
   static constexpr char SYCL_DEVICELIB_METADATA[] = "SYCL/devicelib metadata";
   static constexpr char SYCL_KERNEL_PARAM_OPT_INFO[] = "SYCL/kernel param opt";
   static constexpr char SYCL_PROGRAM_METADATA[] = "SYCL/program metadata";

llvm/include/llvm/Support/PropertySetIO.h

@@ -48,7 +48,7 @@ add_llvm_component_library(LLVMSYCLLowerIR
   SYCLAddOptLevelAttribute.cpp
   SYCLConditionalCallOnDevice.cpp
   SYCLCreateNVVMAnnotations.cpp
-  SYCLDeviceLibReqMask.cpp
+  SYCLDeviceLibBF16.cpp
   SYCLDeviceRequirements.cpp
   SYCLKernelParamOptInfo.cpp
   SYCLJointMatrixTransform.cpp

llvm/lib/SYCLLowerIR/CMakeLists.txt

@@ -0,0 +1,61 @@
+//==----- SYCLDeviceLibBF16.cpp - get SYCL BF16 devicelib required 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
+//
+//===----------------------------------------------------------------------===//
+// This file provides some utils to analyze whether user's device image does
+// depend on sycl bfloat16 device library functions.
+//===----------------------------------------------------------------------===//
+
+#include "llvm/SYCLLowerIR/SYCLDeviceLibBF16.h"
+#include "llvm/ADT/SmallVector.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/Module.h"
+#include "llvm/TargetParser/Triple.h"
+
+static constexpr char DEVICELIB_FUNC_PREFIX[] = "__devicelib_";
+
+using namespace llvm;
+
+static llvm::SmallVector<StringRef, 14> BF16DeviceLibFuncs = {
+    "__devicelib_ConvertFToBF16INTEL",
+    "__devicelib_ConvertBF16ToFINTEL",
+    "__devicelib_ConvertFToBF16INTELVec1",
+    "__devicelib_ConvertBF16ToFINTELVec1",
+    "__devicelib_ConvertFToBF16INTELVec2",
+    "__devicelib_ConvertBF16ToFINTELVec2",
+    "__devicelib_ConvertFToBF16INTELVec3",
+    "__devicelib_ConvertBF16ToFINTELVec3",
+    "__devicelib_ConvertFToBF16INTELVec4",
+    "__devicelib_ConvertBF16ToFINTELVec4",
+    "__devicelib_ConvertFToBF16INTELVec8",
+    "__devicelib_ConvertBF16ToFINTELVec8",
+    "__devicelib_ConvertFToBF16INTELVec16",
+    "__devicelib_ConvertBF16ToFINTELVec16",
+};
+
+bool llvm::isSYCLDeviceLibBF16Used(const Module &M) {
+  if (!Triple(M.getTargetTriple()).isSPIROrSPIRV())
+    return false;
+
+  for (auto Fn : BF16DeviceLibFuncs) {
+    Function *BF16Func = M.getFunction(Fn);
+    if (BF16Func && BF16Func->isDeclaration())
+      return true;
+  }
+
+  return false;
+}
+
+bool llvm::isBF16DeviceLibFuncDecl(const Function &F) {
+  if (!F.isDeclaration() || !F.getName().starts_with(DEVICELIB_FUNC_PREFIX))
+    return false;
+  for (auto BFunc : BF16DeviceLibFuncs) {
+    if (!F.getName().compare(BFunc))
+      return true;
+  }
+
+  return false;
+}