Skip to content

tobias-stadler/compiler

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

56 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

This repository contains an optimizing compiler framework developed from scratch.

Main components:

  • flexible intermediate representation (IR) with support for static single assignment form
  • Pass-based framework for optimization and lowering
  • custom DSL for IR rewriting and backend specification
  • C language frontend
  • RISC-V backend

This project is very much work-in-progress. The primary design goals are simplicity and consistency: a single flexible IR should take the code from the AST to the final assembly code.

Architecture

IR

A function contains an intrusive linked-list of basic blocks, which contains an intrusive linked-list of instructions. Each instruction contains an opcode and an array of operands. Operands can be of many different types. The most important are:

  • SSADef: inline SSA value definition with an SSAType (e.g i32)
  • ExternSSADef: SSA definition outside of an instruction (e.g. functions, basic blocks, stack slots)
  • SSAUse: reference to an SSA definition
  • Reg: virtual or physical register
  • MInt: arbitrary width (<=64 bits) two's complement integer
  • imm32: 32-bit signed immediate

SSA and register operands maintain def-use chains that allow look-up of all related definition and use operands.

DSL

tools/dsl/ contains a compiler for a custom declarative DSL that is used for architecture specification, IR pattern matching and rewriting. DSL code is compiled to C++ code.

IR pattern rewriting

The DSL can be used to match an IR fragment and rewrite it to another IR fragment. Matching is performed in a greedy fashion, larger patterns are tried first. DSL IR patterns closely resemble the textual debug representation of the IR.

Example pattern used during instruction selection:

ir_pat {
  match {
    CONST_INT def(%c,i32) MInt(32,0);
    CMP def(%cmp,i32) eq %lhs %c;
  }
  emit {
    riscv::SLTIU def(%cmp,i32) %lhs imm32(1);
  }
}

More complicated behaviour can be specified using built-in templating facilities and inline C++ code.

Instruction Selection

Instruction selection is performed using the rewriting system explained above. Multiple passes over the IR are performed and different sets of patterns are applied during each pass. Main passes:

  • InstrExpansionPass: canonicalize and legalize instructions to ensure that later the instruction selection patterns will match
  • InstrCombinePass: apply peephole optimization patterns until a fixed-point is reached
  • InstrSelectPass: apply the instruction selection patterns

Register Allocation

The current register allocator is based on a Chaitin-Briggs-style iterated optimistic graph coloring approach.

Main passes:

  • LiveFlowPass: calculate live-in and live-out registers of each block using dataflow analysis
  • LiveGraphPass: construct register interference graph (triangular bitset + adjacency lists)
  • ColoringRegAllocPass: find node elimination order based on degree and spill cost, assign registers, insert spill code
  • RegRewritePass: rewrite virtual registers to the physical registers assigned by the allocator

Current implementation status:

  • COPY instructions are used as hints for guided coloring
  • Spill cost estimation based on the number of register uses and interference graph degree
  • All spilling is global

Before register allocation, SSA form is deconstructed by first restoring conventional SSA-form (by splitting the live-ranges of all PHI defs and uses using COPYs) and then lowering the PHI to a virtual register. To facilitate liveness analysis, all other SSA defs are also lowered to virtual registers.

C frontend

  • Lexer:
    • hand-written two-layer state-machine
  • Preprocessor:
    • fully "pull-based": tokens are preprocessed only on demand
  • Parser:
    • hand-written recursive descent parser
    • operator precedence parsing for expressions
    • fully "LL(1)": only requires Lexer support for next() and peek()
    • no backtracking
    • some integrated semantic analysis:
      • declarations are directly parsed into types
      • symbol table is populated while parsing
    • error recovery: dump source location and parser recursion context
  • IR Generation:
    • lazy SSA construction using approach by M. Braun et al.

The goal is to reliably compile reasonable C99 code, but supporting all features and full conformance is out of scope. Most language constructs are already implemented in Lexer, Preprocessor, Parser and AST, but have not yet been tested extensively (see TODO section for a list of known issues). IR generation currently only supports integer expressions (+ address-of/dereference), structured control flow, function calls with register-sized datatypes, basic structs/unions.

RISC-V backend

  • instruction selector currently supports RV32I instruction set
  • emits textual GNU assembly

Example

int calc(int a, int b) {
  int res = 0;
  if (a > 0) {
    res = a + b;
  }
  return res;
}

Generated IR:

%0:
REF_PARAM def(%3,i32)
REF_PARAM def(%4,i32)
CONST_INT def(%5,i32) MInt(32,0)
CONST_INT def(%6,i32) MInt(32,0)
CMP def(%7,i1) gt %3 %6
BR_COND %7 %1 %2
%1:
ADD def(%8,i32) %3 %4
BR %2
%2:
PHI def(%9,i32) [%8 %1 %5 %0]
RET %9

After instruction selection:

%0:
COPY def(%3,i32) X10
COPY def(%4,i32) X11
COPY def(%5,i32) X0
COPY def(%6,i32) X0
BLT %6 %3 %1 %2
%1:
ADD def(%7,i32) %3 %4
JAL X0 %2
%2:
PHI def(%8,i32) [%7 %1 %5 %0]
COPY def(X10) %8
JALR def(X0) X1 imm32(0) X10

After register allocation and branch lowering:

%0:
ADDI def(X5) X0 imm32(0)
BLT X0 X10 %1
JAL X0 %2
%1:
ADD def(X5) X10 X11
%2:
ADDI def(X10) X5 imm32(0)
JALR def(X0) X1 imm32(0) X10

TODO

C Frontend

  • Lexer:
    • bigraphs and trigraphs
  • Preprocessor:
    • #if expression evaluation (blocked on const-expression evaluation), #pragma, #error, #line
    • correctly disable already expanded macro names
    • token concatination and stringification
  • Parser and AST:
    • support K&R function declaration syntax
    • fix array declarators
    • fix handling of string, integer and char literals
    • fix handling of function prototypes
  • IR generation:
    • blocker: const-expression evaluation
    • ABI for aggregates
    • ...

Register Allocation

  • optimized spilling: live-range splitting, rematerialization (e.g. constants)
  • better spill cost heuristic (loop analysis)
  • dedicated coalescing phase

RISC-V

  • PostRALowering: proper branch relaxation
  • ABILowering: fix stack passing
  • floating point
  • assembler, elf encoder, ...

About

optimizing compiler framework from scratch

Resources

Stars

Watchers

Forks

Languages