Pattern-based Peephole Optimizer

The basic patterns for pattern-based peephole optimizer are collecting from local code analyzer. As result to recursive way, the pattern-based peephole optimizer can cover all optimization cases. This can be an efficient solution for resource conflict situations and need not extra phase to deal with resource conflict. In addition, in the matching process, we can check the modification flag of the IR node to distinguish the node is modified or not. If the node is modified and matched, we can regard it as a new extended pattern. The pattern extends as far as possible will help us to observe the JCS rules widely or the function call applied by the emitters. Using patterns to describe the pattern and matching with recursive way can reduce time to write similar patterns.

discussion (as shown in Figure 6). The first one is pattern matcher, which is generated by the pattern-matcher generator. The internal functions of pattern matcher will be called by pattern-based peephole optimizer and determine whether the instruction sequences is replaceable or not. The pattern matcher is introduced in 3.3.1.

The second part is pattern-matcher generator which is used to generate pattern matcher automatically. This part is introduced in 3.3.2. The third part cost function is implicit in the pattern matcher, which can be used to valuation the benefits of optimizing patterns and to obtain the best optimization steps. This part is introduced in 3.3.3. The latest one is recurrence peephole optimizer which is the main program in pattern-base peephole optimizer. Through the recursive way to finding extended pattern and return to the developers. This part is introduced in 3.3.4.

Figure 6 Framework of pattern-based peephole optimizer.

3.3.1 Pattern Matcher

Pattern matcher is generated by the pattern-matcher generator automatically.

be called by recurrence peephole optimizer. Pattern matcher has three major tasks.

First, match patterns. If we find the matched pattern, we would record it into modification flag. Second, replacement instruction sequences. Once we match the pattern, we would replace instructions with efficient ones. If the instruction has been replaced before, that means, we find new extended pattern in the program. Finally, return the optimization gains. In the end of the function would calculate the benefits of optimization and return the gains. If no matching process invoke, the function would return zero.

3.3.2 Pattern-Matcher Generator

The input of pattern-matcher generator is instruction descriptor which is base on the canonical form [19]. We support three type of description. They are registers descriptor, immediate descriptor and address descriptor. Those descriptors have its own ID with the order of increment number. The immediate descriptor not only supports the constant type (ex.c0, c1), but also the real immediate values (ex. 0, 1).

In the process of generating pattern matcher, if the descriptor position has been defined, we must check the correctness between the current descriptor and defined descriptor. If not, record current position into table.

The first line of pattern descriptor should be total pattern number T, and with T continuous patterns. The format of each pattern is shown below.

z Pattern ID number

z Number of reduced instructions z Number of replaced instructions z Input/Output instruction length

The function name format is the name of the first opcode with a pattern ID number (ex. lwi_0). The position number is the number of each instruction position from zero. If the operand is the register descriptor, we would check the register is defined or not in the table. If defined, we should comparison current node register with defined register and output the comparison code into pattern matcher. If not defined, we would record current position and register type(Rt, Ra or Rb) into table. If the operand is the address descriptor, we check the definition of position in the address table. If defined, we generate the comparison instruction into pattern matcher.

If not defined, we record current position into the address table. If the operand is the constant descriptor, it has the same action as above. In addition, we support the summation function for the constant descriptor.

3.3.3 Cost Function

The concept of the cost function is implicit in the internal pattern matcher.

Calculating the benefits of pattern replacement through the definition of reduce gain and replace gain each time. The calculating value would be used to find the best solution in the recurrence optimizer.

3.3.4 Recurrence Peephole Optimizer

The optimization of recurrence peephole optimizer is based on continuous sequences. We would make a copy before doing optimization so that facilitate to compare the benefit of applying different patterns. Every time matching successfully, we would call recurrence function again till no pattern bas been matched. When the recurrence optimization process has been done, we should record the max gain and return it. The way of using recursive call could achieve the objective of exhaustive matching. The algorithm of recurrence peephole optimizer is shown in Table 14.

Table 14 The algorithm of recurrence peephole optimizer.

Function RPO

Backup Instructions Max_Gain = 0;

Local_Gain =0;

While Current_Node not equal to End_Node For each patterns i

Local_Gain = Pattern_function[i] (Current_Node) if(L_Gain != 0)

L_Gain = L_Gain + RPO End if

Max_Gain = Local_Gain > Max_Gain ? Local_Gain : Max_Gain Local_Gain = 0;

End for

Current_Node = Current_Node ->next_Node;

End while

Restore Instructions return Max_Gain;

End Function