[CMAKE] Upgrade TVM build baseline to C++20

CMakeLists.txt

@@ -439,9 +439,9 @@ include(cmake/utils/CCache.cmake)
 
 include(CheckCXXCompilerFlag)
 set(CMAKE_CXX_STANDARD_REQUIRED ON)
-set(CMAKE_CXX_STANDARD 17)
+set(CMAKE_CXX_STANDARD 20)
 set(CMAKE_CUDA_STANDARD_REQUIRED ON)
-set(CMAKE_CUDA_STANDARD 17)
+set(CMAKE_CUDA_STANDARD 20)
 
 # Module rules
 include(cmake/modules/CUDA.cmake)

cmake/utils/FindLLVM.cmake

@@ -254,9 +254,9 @@ macro(find_llvm use_llvm)
     # compiler-appropriate form so the probe works under MSVC as well.
     if(NOT CMAKE_CXX_STANDARD)
       if(MSVC)
-        set(CMAKE_REQUIRED_FLAGS "/std:c++17")
+        set(CMAKE_REQUIRED_FLAGS "/std:c++20")
       else()
-        set(CMAKE_REQUIRED_FLAGS "-std=c++17")
+        set(CMAKE_REQUIRED_FLAGS "-std=c++20")
       endif()
     endif()
     check_cxx_source_compiles("

docs/install/from_source.rst

@@ -35,11 +35,15 @@ Apache TVM requires the following dependencies:
 - CMake (>= 3.24.0)
 - LLVM (recommended >= 15)
 - Git
-- A recent C++ compiler supporting C++ 17, at the minimum
-    - GCC 7.1
-    - Clang 5.0
-    - Apple Clang 9.3
-    - Visual Studio 2019 (v16.7)
+- A recent C++ compiler supporting C++ 20, at the minimum
+    - GCC 10
+    - Clang 10
+    - Apple Clang 14
+    - Visual Studio 2022
+
+  Optional dependencies that use newer C++20 standard library facilities, such
+  as ``std::format``, may require a newer standard library (for example GCC 13
+  or newer on Linux).
 - Python (>= 3.10)
 - (Optional) Conda (Strongly Recommended)
 

include/tvm/tirx/op.h

@@ -825,7 +825,8 @@ inline bool IsPointerType(const Type& type, const DataType& element_type) {
  * \param span The location of this operation in the source.
  */
 template <typename ValueType,
-          typename = typename std::enable_if<std::is_pod<ValueType>::value>::type>
+          typename = typename std::enable_if<std::is_standard_layout<ValueType>::value &&
+                                             std::is_trivial<ValueType>::value>::type>
 inline PrimExpr make_const(DataType t, ValueType value, Span span = Span());
 /*!
  * \brief Make a const zero expr.

src/relax/transform/run_codegen.cc

@@ -155,7 +155,7 @@ class CodeGenRunner : ExprMutator {
     if (opt_codegen) {
       auto ext_symbol = GetExtSymbol(func);
       size_t count = 0;
-      PostOrderVisit(func->body, [=, &count](Expr e) {
+      PostOrderVisit(func->body, [=, this, &count](Expr e) {
         if (e->IsInstance<ConstantNode>()) {
           // Make sure to pick a unique name
           auto name = ext_symbol + "_" + opt_codegen.value() + "_const_" + std::to_string(count++);

src/runtime/extra/contrib/cudnn/cudnn_json_runtime.cc

@@ -162,7 +162,7 @@ class cuDNNJSONRuntime : public JSONRuntimeBase {
                            conv_dtype, false, &best_algo);
 
     int algo = best_algo.cast<int>();
-    std::function<void()> op_exec = [=]() {
+    std::function<void()> op_exec = [=, this]() {
       int device_id;
       CUDA_CALL(cudaGetDevice(&device_id));
       cudaStream_t stream = static_cast<cudaStream_t>(TVMFFIEnvGetStream(kDLCUDA, device_id));
@@ -223,7 +223,7 @@ class cuDNNJSONRuntime : public JSONRuntimeBase {
     auto runner = tvm::contrib::CuDNNSDPARunner::Create();
     runner->Init(batch, seq_len, num_heads, num_kv_heads, head_size, head_size_v, scale, dtype,
                  layout);
-    return [=]() {
+    return [=, this]() {
       auto qkv = GetInput(node, 0);
       auto workspace = const_cast<DLTensor*>(GetInput(node, 1));
       auto out = const_cast<DLTensor*>(data_entry_[EntryID(outputs_[0])]);

src/runtime/extra/disco/protocol.h

@@ -28,6 +28,7 @@
 
 #include <memory>
 #include <string>
+#include <type_traits>
 #include <utility>
 #include <vector>
 
@@ -78,7 +79,8 @@ struct DiscoProtocol {
   /*!\ brief Arena used by RPCReference to allocate POD memory */
   template <typename T>
   T* ArenaAlloc(int count) {
-    static_assert(std::is_pod<T>::value, "need to be trival");
+    static_assert(std::is_standard_layout<T>::value && std::is_trivial<T>::value,
+                  "need to be trivial");
     return arena_.template allocate_<T>(count);
   }
 

src/runtime/rpc/rpc_endpoint.cc

@@ -34,6 +34,7 @@
 #include <cmath>
 #include <memory>
 #include <string>
+#include <type_traits>
 #include <utility>
 #include <vector>
 
@@ -312,7 +313,8 @@ class RPCEndpoint::EventHandler : public support::Stream {
 
   template <typename T>
   T* ArenaAlloc(int count) {
-    static_assert(std::is_pod<T>::value, "need to be trival");
+    static_assert(std::is_standard_layout<T>::value && std::is_trivial<T>::value,
+                  "need to be trivial");
     return arena_.template allocate_<T>(count);
   }
 

src/s_tir/transform/compact_buffer_region.cc

@@ -129,6 +129,8 @@ class BufferAccessRegionCollector : public StmtExprVisitor {
   }
 
  private:
+  using StmtExprVisitor::VisitBufferDef;
+
   struct BufferAccessInfo {
     /*! \brief The buffer. */
     Buffer buffer;

src/s_tir/transform/inject_software_pipeline.cc

@@ -707,7 +707,7 @@ class PipelineRewriter : public StmtExprMutator {
         }
       }
 
-      auto wait_count = [=, &ana_normalized]() {
+      auto wait_count = [=, this, &ana_normalized]() {
         auto sum = PrimExpr(0);
         for (auto producer_head : producer_head_per_commit) {
           if (producer_head && ana_normalized->CanProve(producer_head.value() >= 0)) {

src/target/llvm/codegen_params.cc

@@ -61,7 +61,8 @@ struct LLVMConstantGetter<T, std::enable_if_t<std::is_floating_point<T>::value>>
   static llvm::Constant* getElement(llvm::Type* ty, T t) { return llvm::ConstantFP::get(ty, t); }
 };
 
-template <typename T, typename = std::enable_if<std::is_pod<T>::value>>
+template <typename T,
+          typename = std::enable_if_t<std::is_standard_layout<T>::value && std::is_trivial<T>::value>>
 void BuildLLVMVector(llvm::Type* element_type, void* tensor_data, size_t num_elements,
                      std::vector<llvm::Constant*>* elements) {
   elements->resize(num_elements, nullptr);