Support LLVM backend #171

leewei05 · 2024-07-01T12:36:31Z

Passed all code generation tests. 🥳

Changes of the old infrastructure

Testing: I use lli for testing generated LLVM IR code. I assumed that lli can link printf to libc, which is it's own c library, and it actually worked! In order to make this work, I did a little bit of hacking in LLVMIRGenerator::FuncCallExprNode. Whenever, there is a __builtin_print function being called, it will replaced it with printf, so that we don't need to modify our test files.
Makefile: I added a few flags with llvm-config.
main.cpp: separated QBE and LLVM code generation process with two helper functions. Added LLVM as target option, maybe we can consider changing LLVM as default target. 🐉
Type: created fields accessor for record type. Added MemberIndex for record type for getting the index number of members. Union will always return index 0 if member is found. Lastly, changed return_type() of FuncType to support LLVMIRUtil::GetLLVMType(), which may also need additional help.

Newly introduced

include/llvm/util.hpp, src/llvm/util.cpp: Some utility functions for creating LLVM IR. Also, some suggestions for storing builder_ are appreciated! 🙏
- GetLLVMType function translate our Type into corresponding llvm::Type*, which can be directly used for CreateCall/Load/Store etc.
include/llvm_ir_generator.hpp: Initialize some LLVM IR builder objects. I saw people say that LLVMContext is for LLVM developers to know the details. Right now, we only need to know LLVMContext includes LLVM core infrastructures. Module is kind of like translation unit in our parser, it knows functions, global variables and symbol table, which I didn't use (Maybe it has some handy functions we can use). IRBuilder is the key object for using LLVM API to construct LLVM IR.
src/llvm_ir_generator.cpp: The main LLVM IR code generator file. Some differences worth noticing:
- Comparison operator, such as, greater than, equal than, are not Binary operators in LLVM. It has a different class called llvm::CmpInst, so the implementation is a bit different than QBE. I think they separated them because comparison may have different low-level computer architecture optimization techniques than other binary operations.
- I stole some of the ideas from QBE generator, such as num_recorder, it is val_recorder in LLVM IR generator instead. Since llvm::Value is a super important class, and any class can literally be llvm::Value class, except for llvm::Type, I decided to store llvm::Value* in the val_recorder. Upper level node can directly use the value stored in it.
- We use GetElementPtr for locating memory address in Array, Struct, Union.
- Each builder_->Create... will create a dest number that we can use.
- Almost every builder_->CreateAlloc/CreateCall/CreateStore/CreateLoad requires the callee's type. It can be tricky for function pointers because we cannot get the element type of pointer. The reason is to support Opaque Pointers. The solution for function pointer's type is to return llvm::FunctionType instead of llvm::PointerType for GetLLVMType.
- Every Basic block must only have one terminator instruction, such as br, ret. llvm::BasicBlock* can also represent as label. If we want to jump to label xx, we would jump to basic block xx.

Help wanted!

No help needed! I refactored GetLLVMType and remove redundant code! 😄

As I mentioned that calling a function requires llvm::FunctionType* ( return type + parameter types), function pointers can be tricky to handle. The ideal way is to get LLVM type from our current type system, which is the job of LLVMIRUtil::GetLLVMType, it can return the correct LLVM type for a given Type, similar with your ResolvedType. Right now, I have a val_to_type map for mapping llvm::Value* to llvm::Type*. This way, I can get the correct function type of the corresponding llvm::Value* to make a function call. My plan is to get rid of val_to_type and fully depend on our type system to get the correct LLVM type.

The current bottleneck is to change PtrType's base_type() from const Type& base_type() to const std::unique_ptr<Type>& base_type(), so that I can pass the base type of PtrType to LLVMIRUtil::GetLLVMType. Changing this also require to change IsEqual, IsConvertible etc.

Lai-YT

Overall, this is a great PR that makes significant progress, and I enjoyed reviewing it. As I am not familiar with the LLVM APIs, there's not much I can support. However, as all test cases have passed, I believe there are no major issues. Small ones can be found and addressed as we evolve in the future. 🔥

I have submitted some minor revisions that I hope will make the PR even better. 😄

main.cpp

src/llvm/util.cpp

include/llvm/util.hpp

include/llvm_ir_generator.hpp

src/llvm_ir_generator.cpp

include/llvm/util.hpp

include/llvm_ir_generator.hpp

Lai-YT · 2024-07-05T12:41:00Z

include/llvm_ir_generator.hpp

+  /// @brief A LLVM object that includes core LLVM infrastructure.
+  std::unique_ptr<llvm::LLVMContext> context_;
+  /// @brief Provides LLVM Builder API for constructing IR. By default, Constant
+  /// folding is enabled and we have more flexibility for inserting
+  /// instructions.
+  std::unique_ptr<llvm::IRBuilder<>> builder_;
+  /// @brief Stores global variables, function lists, and the constructed IR.
+  std::unique_ptr<llvm::Module> module_;


They seem to be passed by reference when used, meaning they are not going to be copied. Furthermore, I believe these classes are not copyable, thus already unique within the code generator.

main.cpp

src/llvm_ir_generator.cpp

Lai-YT · 2024-07-05T12:47:28Z

src/llvm_ir_generator.cpp

+  auto var_type = llvm_util_.GetLLVMType(*(decl.type));
+  // For function pointer, we need to change from FunctionType to PointerType
+  var_type = var_type->isFunctionTy() ? var_type->getPointerTo() : var_type;


Got it. I didn't realize the type is already modified by the conversion.

include/llvm_ir_generator.hpp

Lai-YT · 2024-07-05T13:20:15Z

src/llvm_ir_generator.cpp

+    }
+  } else if (val->getType()->isPointerTy()) {
+    // function pointer
+    auto type = llvm_util_.GetLLVMType(*(call_expr.func_expr->type));


When will LLVM require a function pointer to have a pointer type? I'm considering having GetLLVMType return the pointer type for function pointers directly. Would that be more straightforward?

Lai-YT · 2024-07-05T13:25:35Z

src/llvm_ir_generator.cpp

+  if (id_expr.type->IsPtr() || id_expr.type->IsFunc()) {
+    auto res = builder_->CreateLoad(llvm_util_.IntPtrType(), id_val);
+    val_recorder.Record(res);
+    val_to_id_addr[res] = id_val;
+  } else {
+    auto res =
+        builder_->CreateLoad(llvm_util_.GetLLVMType(*(id_expr.type)), id_val);
+    val_recorder.Record(res);
+    val_to_id_addr[res] = id_val;
+  }


Oh I see. This is because GetLLVMType returns FunctionType on function pointers. 😅

Lai-YT · 2024-07-05T13:27:04Z

src/llvm_ir_generator.cpp

+  auto res_addr = builder_->CreateConstInBoundsGEP2_32(
+      arr_type, base_addr, 0, (unsigned int)index->val);


Wow, there's a dedicated page for GEP. 😮 It's a complex topic!

Lai-YT

LGTM 🎉

leewei05 added 30 commits June 24, 2024 16:25

Set up initial testing configuration for llvm

e5de4b1

Create llvm IR generator

490766b

Write LLVM IR to files

c312c58

Generate return statement and constant expression LLVM IR

3e90edc

Integrate lli for testing LLVM IR

ef1879c

Update turnt config file and printf with parameter

eca3d05

Handle simple binary expression

fcc5838

Update TransUnitNode and ExternDeclNode

2975b36

Insert __builtin_print_format global variable

c3603aa

Print output with llvm interpreter

9e85f46

Simplify IntConstExprNode by returning integer const directly

c903cc9

Code generation for variable declaration and id expression

df0f59a

Refactor i32Ty with Util class

2d46738

Binary and comparison operator code generation

af47ebd

Implement simple assignment expression code generation

ada573d

Implement prefix increment and decrement code generation

690db76

Implement logical AND and OR operation code generation

af5678e

Implement Bitwise complement operator code generation

b4a9c16

Implement postfix increment and decrement operator code generation

4b250f1

Implement if else statement code generation

e983617

Implement while and do while statement code generation

7ccc800

Implement for statement code generation

bc5c7a5

Implement break and continue statement code generation

6bfb6d0

Rename integer type name

65b7af0

Implement pointer code generation

563fba5

Remove hardcode pointer size

fc9f7d4

Implement function call expression code generation

158ac6e

Implement condition expression code generation

07706ec

Implement array code generation

c93394b

Implement function pointer code generation

fd78532

Lai-YT requested changes Jul 5, 2024

View reviewed changes

leewei05 added 13 commits July 5, 2024 14:27

Rename helper functions

3099ea2

Rename builder helper class and builder reference

4e1b634

Inline builder integer and pointer type

f598de2

Refactor helper function for basic blocks

1f8ca0a

Include missing headers

aa98013

Throw errors for operator helper functions

e22a126

Rename LabelInfo struct and vector

89b08d1

Update comments and remove redundant arguments

6b89405

Refactor create basic block if not exist logic

49d3ab0

Refactor IdExprNode

93dd326

Record return value of printf function

cd17d35

Rename variable

bc6e894

Refactor with GetLLVMType

015682e

leewei05 force-pushed the llvm-backend branch from d5d826b to 17811be Compare July 5, 2024 08:53

Fix clang-tidy

3fa5700

leewei05 force-pushed the llvm-backend branch from 17811be to 3fa5700 Compare July 5, 2024 09:01

leewei05 requested a review from Lai-YT July 5, 2024 09:05

Lai-YT reviewed Jul 5, 2024

View reviewed changes

leewei05 added 3 commits July 5, 2024 22:34

Move HasTerminator outside of BuilderHelper class

a2ce9e0

Rename builder_helper_ member and update comments

9c0088f

Pass constant string reference

6e50cce

leewei05 requested a review from Lai-YT July 5, 2024 15:17

Refactor LLVM IR generator data members

c030086

leewei05 force-pushed the llvm-backend branch from 01bd432 to a857a9d Compare July 5, 2024 15:20

Lai-YT approved these changes Jul 5, 2024

View reviewed changes

Update README

f212231

leewei05 force-pushed the llvm-backend branch from a857a9d to f212231 Compare July 5, 2024 15:26

Lai-YT merged commit d787d64 into fruits-lab:main Jul 5, 2024
4 checks passed

leewei05 deleted the llvm-backend branch July 5, 2024 15:28

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Support LLVM backend #171

Support LLVM backend #171

leewei05 commented Jul 1, 2024 •

edited

Loading

Lai-YT left a comment

Lai-YT Jul 5, 2024

Lai-YT Jul 5, 2024

Lai-YT Jul 5, 2024

Lai-YT Jul 5, 2024

Lai-YT Jul 5, 2024

Lai-YT left a comment

		auto res_addr = builder_->CreateConstInBoundsGEP2_32(
		arr_type, base_addr, 0, (unsigned int)index->val);

Support LLVM backend #171

Support LLVM backend #171

Conversation

leewei05 commented Jul 1, 2024 • edited Loading

Changes of the old infrastructure

Newly introduced

Help wanted!

Lai-YT left a comment

Choose a reason for hiding this comment

Lai-YT Jul 5, 2024

Choose a reason for hiding this comment

Lai-YT Jul 5, 2024

Choose a reason for hiding this comment

Lai-YT Jul 5, 2024

Choose a reason for hiding this comment

Lai-YT Jul 5, 2024

Choose a reason for hiding this comment

Lai-YT Jul 5, 2024

Choose a reason for hiding this comment

Lai-YT left a comment

Choose a reason for hiding this comment

leewei05 commented Jul 1, 2024 •

edited

Loading