Check out the new USENIX Web site. next up previous
Next: Other VM Subsystems Up: Establishing functionality Previous: Baseline Compiler

Optimizing Compiler

Because writing an optimizing compiler represents a significant investment, from the beginning we designed the optimizing compiler for portability. A key aspect of this design divides the intermediate representation (IR) into three levels of operators:
  1. High Level IR (HIR): The HIR operator set is architecture independent and resembles the bytecode instruction set, although the IR uses a register transfer language in place of the bytecode stack abstraction.
  2. Low Level IR (LIR): The LIR operator set is architecture independent and resembles the instruction set of a typical RISC machine. The main difference between HIR and LIR is that complex HIR operators such as new or call virtual are expanded into the appropriate sequence of primitive operations.
  3. Machine Level IR (MIR): The MIR operator set is architecture-specific; with the exception of a few pseudo-operators that are expanded as part of final assembly, the MIR provides a one-to-one mapping between operators and the target ISA.
Translation from bytecodes to HIR, HIR optimizations, translation from HIR to LIR, and LIR optimizations are all architecture independent.

At system build-time, the builder generates classes representing the IR operators and helper functions from template files. Thus, the task of defining MIR operators corresponding to IA32 instructions consisted of defining new instruction formats and operator characteristics in two machine-dependent template files. The MIR operator definitions for IA32 consist of about 1300 lines in two files, defining 153 operators falling into 29 formats.

After defining the MIR, the main tasks in porting the optimizing compiler to a new ISA involve implementing the three major architecture-specific compiler phases: translation from LIR to MIR (aka instruction selection), register allocation, and final assembly. Each of these stages actually contains both architecture-independent and architecture-specific portions; we adopted a common idiom to define abstract classes that implement the shared functionality in terms of abstract methods defined by architecture-specific subclasses.

The instruction selection phase translates the machine-independent LIR into architecture-specific MIR. This translation phase partitions the dependence graph of each basic block into a forest of trees, and feeds the forest to a Bottom-Up Rewrite System (BURS)-based tree-pattern matching system [7]. Thus, the major porting task is to define the tree patterns and associated actions that serve as the input grammar to the BURS engine.

The register allocator maps the infinite set of symbolic registers onto a finite set of physical registers and spill locations. In addition, this phase generates prologues, epilogues, and calling sequences that respect the Jikes RVM and/or native OS calling conventions. The optimizing compiler relies on a variant of linear-scan [9] register allocation. The core of the register allocator should be machine-independent, but the implementation handles a fair number of low-level architecture-specific issues. Unfortunately, the original PowerPC implementation deeply intertwined the machine-dependent and machine-independent register allocator code. Rather than tackle a major refactoring problem with the old implementation, we decided to re-implement the register allocator from scratch for IA32. The new implementation cleanly separates machine-dependent and machine-independent code and includes expanded functionality and heuristics to suit both register-scarce and register-rich architectures. We have since back-ported the new implementation to PowerPC, so both platforms now share the machine-independent portion of the register allocator.

Final assembly generates executable machine code from the MIR and finalizes descriptive data structures such as exception tables and GC maps. The code for generating the descriptive data structures does not depend on the target architecture. Furthermore, as described above, the build system mechanically generates code that interfaces the MIR to the assembler. Therefore this stage introduced little work for porting.

The optimizing compiler also performs some optimizations on the MIR. The main MIR peephole optimizations, branch simplification and null check folding, do not depend on the architecture and worked immediately on IA32. We have not yet ported the instruction scheduler to IA32.

Figure 1: Examples of architectural register restrictions in the optimizing compiler IR. a) Java source code; b) IR fragment before instruction selection; c) IR fragment before register allocation; d) IR fragment after register allocation.
byte foo(byte[] a, 
int x, int y){
 int i=x<<y;
 return a[i];
t1 = shl_acc t2 
t3 = byte_aload t0,t1
return t3
ecx = t2
t1 = shl_acc ecx
t3 = movsx [t0+t1]
eax = t3
return eax
ecx = [esp + 12]
edx = shl_acc ecx
eax = movsx [eax+edx]
return eax

One difficulty in generating IA32 code compared to PowerPC is dealing with register restrictions imposed by the non-orthogonal instruction set architecture. Figure 1 shows some examples.

For some IA32 instructions, a particular operand must reside in a distinguished register. For example, in Figure 1b, the value of t2 must reside in ecx at the shift instruction (shl_acc). We represent this information in the IR by explicitly assigning t2 to ecx before the shift instruction during instruction selection for the shift (Figure 1c). The register allocator respects liveness for physical registers, and will further attempt to allocate t2 to ecx to remove the copy operation. Similarly, in our calling convention eax holds the return value; we enforces this by inserting a copy from the symbolic return value register to eax, as shown in Figure 1c.

When BURS inserts a memory operand, the register allocator must respect further restrictions. For example, if the allocator were to spill t0 in Figure 1c, this would force a later pass to move t0 to a scratch register before its use in the movsx memory operand. For this reason, the linear scan spill heuristics consider a spill of a symbolic register used in a memory operand to be more expensive than a spill of a register operand that could be replaced by a memory operand representing the spill location.

Furthermore, for many instructions, the IA32 architecture dictates that only four of the eight general-purpose registers can hold 8-bit values. For example, in Figure 1c, t3 can reside in the low word of eax, ebx, ecx, or edx; but not, for example in ebp or edi. The register allocator handles these types of restrictions with special case code, computing the restrictions as a pre-pass to register allocation.

next up previous
Next: Other VM Subsystems Up: Establishing functionality Previous: Baseline Compiler
Stephen Fink 2002-05-23