deep

a Cross Development Platform for Java

User Tools

Site Tools


dev:crosscompiler:backend:register_allocator

Register Allocation

Starting with the SSA form for each result of an SSA instruction a register has to be assigned. However, it must be decided how the registers of a processor will be used generally.

Register Usage

Registers can be classified as dedicated, volatile or nonvolatile. Dedicated are register with a special purpose, such as the stackpointer. Volatile means, that this register might be used any time. It does not have to be stored when another method is called. Nonvolatile registers must be saved in the prologue of a method before they can be used and after use they must be restored. The number of volatile and nonvolatile registers must be equal for all methods.

Resolving of Phi-Functions

The operands of a phi-function must all get the same register as the result of the phi-function. For this purpose all these operands have a field join.
Before the phi-functions can be resolved, loops need special treatment. Let's consider the folowing case:

Phi-Function in a loop

The SSAValue of a is used in the loop. The live range of a extends to the instruction for b = 2 * a. However, this would be wrong, because the register for a would be released at this point and might be used otherwise. It must stay reserved until the end of the loop. At the beginning of node 2 a phi-function is created for a. This function will be deleted because a is set in node 1 and only read in node 2. For all phi-functions in loops (whether deleted or not) the field last is set to the last instruction of this loop.
Subsequently, all instructions of all SSA nodes are copied into an array, since the control flow is no longer important for the further processing and the access is faster.
All used phi-functions are marked (field used). used indicates if a deleted phi-Funktion is used further downstream. If not, the register can be released immediately. The reason for doing this is the fact, that phi-functions were inserted for every local variable if loops are present. Most of this phi-functions are deleted and should not occupy a register unnecessarily.
Phi-functions will be resolved as follows: A new state array join is created. The field join of the result of the phi-function and of its operands refer to the value in this state array with the same index as the phi-function. Several phi-functions might refer to the same join value.
There are cases where two different phi-functions with the same index refer to two different variables, as in:

  if() {
    int a = ...;
    while() a = ...;
  } else {
    float f = ...;
    while() f = ...;
  }

a and f occupy the same slot (or index). Nevertheless, the two ranges must not be integrated into one. If the two variables have the same type it wouldn't really matter, though it wouldn't be efficient. However, in this case the type is different and later, we have to assign different registers. Therefore, we check if the range of a new phi-function overlaps with the range of a phi-function further upstream at the same index. If this is not the case, we create another join value in the state array and link it to the previous one.

Phi-functions for same slots but different types

Determination of Live Range

We determine for the results of all SSA instructions up to which instruction they are used (field end). Results of phi-functions must be handled separately. Such results might be used as operands of instructions which are placed further upstream. For this reason the live range of a phi-function does not start with its instruction but with its first use of an operand or the phi-function itself. Hence, phi-function need a field start.

Determine the Register Type

Before beeing able to assign registers we have to decide whether to use a volatile or nonvolatile register. As soon there is a call to a method within the live range of a value (holds also for phi-functions) then this value must be assigned a nonvolatile Register. Other values are assigned a volatile register. The same is true for exceptions. If the live range of a variable extends into (or over) a catch block we must assign a nonvolatile register, as the handling of the exception destroys volatiles.
A special case is the SSA instruction sCloadLocal. With this we must ckeck if a method call occurs between the start of the method to the end of the live range. If this is not the case we can leave the parameter in the the register where it was passed, which is a volatile register already.

Determine Used Parameters

Parameters are passed in volatile registers. Some of them have to be copied to nonvolatile registers. Some parameters might not be used at all in a method. This can be determined in the exit set of the last node. If a value is null in that set the parameter with this index was never loaded and hence no register must be reserved. The following example demonstrates parameter usage on a PPC platform.

  float m2(long a, float b, double c, byte[] d, short e, int f, int g) {
    a = 0x7545 & a;
    e += 100;
    T08Calls.classMethod(d[2]);
    e = (short)(20 + e);
    g |= 0x223344;
    c = 3.2;
    int h = g - e;
    T08Calls.classMethod(h);
    short i = (short)h;		
    return b;
  }
Variablethisabcdefg
TyprefLongFloatDoubleByteArrayShortIntegerInteger
Register Typ volnonVolvolvolvol nonVol
not usedx x
Register R3,R4FR31FR2R5R6 R31

Register Assignment

Finally the results of all SSA instructions will be assigned a register.

Releasing of Registers of phi-Functions

As described above, phi-functions have a field last which is used to make sure that a registers of a phi-function lives to the very end of its loop. At the this end the register of the phi-function (which is actually given in its join value) could be released. However, at this stage we already passed the phi-function. For this purpose all phi-functions are referenced in a linked list (with the field next) in the first SSA instruction of the node following the node with the phi-function.

Loading of Parameters

All sCloadLocal instructions use the same register as their parameter. In case of join != null, the referenced phi function gets the same register.

Reservation of Auxiliary Registers

Some instruction need 1 or 2 additional (auxiliary) registers. These will be reserved and stored in regAux1 and regAux2. At the end of the instruction they can be released again. It's possible that the result of the SSA instruction occupies the same register as one of the auxiliary registers.

Releasing of Operand Registers

Registers of operands, whose live range expires, must be released. If they are of type long, this step must happen at the very end.

Reservation of a Result Register

The register pool will be searched for a free register for the result. The result type determines if 1 or 2 GPR's or a FPR must be reserved. The following two cases must be considered.

1. join != null

The same register must be assigned as the result of the phi function. If not handled yet the phi function has to be dealt with first. It can happen that several phi funtions occupy the same index in the state array:

  switch(i) {
  case 0: return 0;
  case 1: i++; break;
  case 2: return 2;
  case 3: i += 3;
  case 4: i += 4; break;
  default: return -1;
  ..
  return i + 3;

In those cases all phi functions with the same index get the same register assigned.

2. Loading of a Constant

If a constant can be used as immediate operand by the machine instruction no register has to be reserved. However, if the value shows up in the state array (index >= 0), it must be loaded into a register anyway. This is the case with the conditional operator. A register must be assigned also in case of join != null.

Releasing of Result Registers

The result register can be released immediately if the live range already expires.

Locals on the Stack

Local variables which cannot be assigned a register get assigned a stack slot. Stack slots are numbered from 0x100. The code generator will handle the transfer to and from these stack slots. Whenever the code generator tries to translate an SSA instruction and one of the operands is on the stack, the following has to be dones:

  • load stack slot into free register
  • execute instruction

If the result of the SSA instruction is assigned a stack slot, the code generator will have to save to the stack:

  • execute instruction, destination register is free register
  • store destination register onto stack

During register allocation of a method a variable fullRegSet signals, if at least one stack slot was used. If that was the case the register allocation is done again, but this time with a reduced register set, because some of the registers must be free in order to be used for fetching stack slots and using them in the following instructions.

dev/crosscompiler/backend/register_allocator.txt · Last modified: 2019/08/29 16:53 by ursgraf