[llvm] Drop code for llvm 15. (taichi-dev#5313)

* Drop code for llvm 15

Still use clang10 for COMPILE_LLVM_RUNTIME. Most changes are related to the opaque ptr change in llvm 15 which requires type when create Load and GEP. It would be nice to generate the type from taichi type.

Known issue.
Still some getPointerElementType use which is Deprecated for llvm 15.
Also some cases use hack to get llvm type for Load/GEP.

Only tested python tests/run_tests.py -v -t3 -a cpu -s on windows.
2 crashes
crashes:
  fp16 crash when hw not support fp16 for vulkan.
  element_wise crashed in llvm pass, need more time to debug.

CMakeLists.txt

@@ -80,7 +80,10 @@ if (WIN32)
   #
   # FIXME: (penguinliong) This is fixed in later releases of LLVM so maybe
   # someday we can distribute `Debug` libraries, if it's ever needed.
-  SET(CMAKE_MSVC_RUNTIME_LIBRARY MultiThreadedDLL)
+  if (NOT TI_LLVM_15)
+    SET(CMAKE_MSVC_RUNTIME_LIBRARY MultiThreadedDLL)
+  endif()
+  message("CMAKE_MSVC_RUNTIME_LIBRARY: ${CMAKE_MSVC_RUNTIME_LIBRARY}")
 endif()
 
 # No support of Python for Android build; or in any case taichi is integrated

cmake/TaichiCore.cmake

@@ -1,5 +1,6 @@
 option(USE_STDCPP "Use -stdlib=libc++" OFF)
 option(TI_WITH_LLVM "Build with LLVM backends" ON)
+option(TI_LLVM_15 "Switch to LLVM 15" OFF)
 option(TI_WITH_METAL "Build with the Metal backend" ON)
 option(TI_WITH_CUDA "Build with the CUDA backend" ON)
 option(TI_WITH_CUDA_TOOLKIT "Build with the CUDA toolkit" OFF)
@@ -142,6 +143,10 @@ else()
     list(REMOVE_ITEM TAICHI_CORE_SOURCE ${TAICHI_LLVM_SOURCE})
 endif()
 
+if (TI_LLVM_15)
+    set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DTI_LLVM_15")
+endif()
+
 if (TI_WITH_CUDA)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DTI_WITH_CUDA")
   file(GLOB TAICHI_CUDA_RUNTIME_SOURCE "taichi/runtime/cuda/runtime.cpp")

taichi/codegen/cpu/codegen_cpu.cpp

@@ -107,8 +107,14 @@ class CodeGenLLVMCPU : public CodeGenLLVM {
 
       {
         builder->SetInsertPoint(loop_test_bb);
+#ifdef TI_LLVM_15
+        auto *loop_index_load =
+            builder->CreateLoad(builder->getInt32Ty(), loop_index);
+#else
+        auto *loop_index_load = builder->CreateLoad(loop_index);
+#endif
         auto cond = builder->CreateICmp(
-            llvm::CmpInst::Predicate::ICMP_SLT, builder->CreateLoad(loop_index),
+            llvm::CmpInst::Predicate::ICMP_SLT, loop_index_load,
             llvm_val[stmt->owned_num_local.find(stmt->major_from_type)
                          ->second]);
         builder->CreateCondBr(cond, loop_body_bb, func_exit);
@@ -121,9 +127,15 @@ class CodeGenLLVMCPU : public CodeGenLLVM {
           auto &s = stmt->body->statements[i];
           s->accept(this);
         }
-        builder->CreateStore(builder->CreateAdd(builder->CreateLoad(loop_index),
-                                                tlctx->get_constant(1)),
-                             loop_index);
+#ifdef TI_LLVM_15
+        auto *loop_index_load =
+            builder->CreateLoad(builder->getInt32Ty(), loop_index);
+#else
+        auto *loop_index_load = builder->CreateLoad(loop_index);
+#endif
+        builder->CreateStore(
+            builder->CreateAdd(loop_index_load, tlctx->get_constant(1)),
+            loop_index);
         builder->CreateBr(loop_test_bb);
         builder->SetInsertPoint(func_exit);
       }

taichi/codegen/cuda/codegen_cuda.cpp

@@ -65,6 +65,9 @@ class CodeGenLLVMCUDA : public CodeGenLLVM {
     auto value_arr = builder->CreateAlloca(stype);
     for (int i = 0; i < values.size(); i++) {
       auto value_ptr = builder->CreateGEP(
+#ifdef TI_LLVM_15
+          stype,
+#endif
           value_arr, {tlctx->get_constant(0), tlctx->get_constant(i)});
       builder->CreateStore(values[i], value_ptr);
     }
@@ -324,8 +327,11 @@ class CodeGenLLVMCUDA : public CodeGenLLVM {
 
     // Use the value from the memory that atomicCAS operates on to initialize
     // cas_old_output.
-    llvm::Value *cas_old_output =
-        builder->CreateLoad(atomic_memory_address, "cas_old_output");
+    llvm::Value *cas_old_output = builder->CreateLoad(
+#ifdef TI_LLVM_15
+        atomic_type,
+#endif
+        atomic_memory_address, "cas_old_output");
     builder->CreateStore(cas_old_output, cas_old_output_address);
 
     llvm::BasicBlock *loop_body_bb =
@@ -338,21 +344,35 @@ class CodeGenLLVMCUDA : public CodeGenLLVM {
     // loop body for one atomicCAS
     {
       // Use cas_old_output to initialize cas_new_output.
-      cas_old_output =
-          builder->CreateLoad(cas_old_output_address, "cas_old_output");
+      cas_old_output = builder->CreateLoad(
+#ifdef TI_LLVM_15
+          atomic_type,
+#endif
+          cas_old_output_address, "cas_old_output");
       builder->CreateStore(cas_old_output, cas_new_output_address);
 
-      auto binop_output = op(builder->CreateLoad(binop_output_address), val);
+      auto binop_output = op(builder->CreateLoad(
+#ifdef TI_LLVM_15
+                                 atomic_type,
+#endif
+                                 binop_output_address),
+                             val);
       builder->CreateStore(binop_output, binop_output_address);
 
-      llvm::Value *cas_new_output =
-          builder->CreateLoad(cas_new_output_address, "cas_new_output");
+      llvm::Value *cas_new_output = builder->CreateLoad(
+#ifdef TI_LLVM_15
+          atomic_type,
+#endif
+          cas_new_output_address, "cas_new_output");
 
       // Emit code to perform the atomicCAS operation
       // (cas_old_output, success) = atomicCAS(memory_address, cas_old_output,
       //                                       cas_new_output);
       llvm::Value *ret_value = builder->CreateAtomicCmpXchg(
           atomic_memory_address, cas_old_output, cas_new_output,
+#ifdef TI_LLVM_15
+          llvm::MaybeAlign(0),
+#endif
           llvm::AtomicOrdering::SequentiallyConsistent,
           llvm::AtomicOrdering::SequentiallyConsistent);
 
@@ -427,8 +447,8 @@ class CodeGenLLVMCUDA : public CodeGenLLVM {
           llvm::BasicBlock::Create(*llvm_context, "loop_body", func);
       auto func_exit =
           llvm::BasicBlock::Create(*llvm_context, "func_exit", func);
-      auto loop_index =
-          create_entry_block_alloca(llvm::Type::getInt32Ty(*llvm_context));
+      auto i32_ty = llvm::Type::getInt32Ty(*llvm_context);
+      auto loop_index = create_entry_block_alloca(i32_ty);
       llvm::Value *thread_idx =
           builder->CreateIntrinsic(Intrinsic::nvvm_read_ptx_sreg_tid_x, {}, {});
       llvm::Value *block_dim = builder->CreateIntrinsic(
@@ -439,7 +459,12 @@ class CodeGenLLVMCUDA : public CodeGenLLVM {
       {
         builder->SetInsertPoint(loop_test_bb);
         auto cond = builder->CreateICmp(
-            llvm::CmpInst::Predicate::ICMP_SLT, builder->CreateLoad(loop_index),
+            llvm::CmpInst::Predicate::ICMP_SLT,
+            builder->CreateLoad(
+#ifdef TI_LLVM_15
+                i32_ty,
+#endif
+                loop_index),
             llvm_val[stmt->owned_num_local.find(stmt->major_from_type)
                          ->second]);
         builder->CreateCondBr(cond, loop_body_bb, func_exit);
@@ -452,9 +477,13 @@ class CodeGenLLVMCUDA : public CodeGenLLVM {
           auto &s = stmt->body->statements[i];
           s->accept(this);
         }
-        builder->CreateStore(
-            builder->CreateAdd(builder->CreateLoad(loop_index), block_dim),
-            loop_index);
+        builder->CreateStore(builder->CreateAdd(builder->CreateLoad(
+#ifdef TI_LLVM_15
+                                                    i32_ty,
+#endif
+                                                    loop_index),
+                                                block_dim),
+                             loop_index);
         builder->CreateBr(loop_test_bb);
         builder->SetInsertPoint(func_exit);
       }