1 | \section{Our algorithm} |
---|
2 | |
---|
3 | \subsection{Design decisions} |
---|
4 | |
---|
5 | Given the NP-completeness of the problem, finding optimal solutions |
---|
6 | (using, for example, a constraint solver) can potentially be very costly. |
---|
7 | |
---|
8 | The SDCC compiler~\cite{SDCC2011}, which has a backend targeting the MCS-51 |
---|
9 | instruction set, simply encodes every branch instruction as a long jump |
---|
10 | without taking the distance into account. While certainly correct (the long |
---|
11 | jump can reach any destination in memory) and a very fast solution to compute, |
---|
12 | it results in a less than optimal solution in terms of output size and |
---|
13 | execution time. |
---|
14 | |
---|
15 | On the other hand, the {\tt gcc} compiler suite, while compiling |
---|
16 | C on the x86 architecture, uses a greatest fix point algorithm. In other words, |
---|
17 | it starts with all branch instructions encoded as the largest jumps |
---|
18 | available, and then tries to reduce the size of branch instructions as much as |
---|
19 | possible. |
---|
20 | |
---|
21 | Such an algorithm has the advantage that any intermediate result it returns |
---|
22 | is correct: the solution where every branch instruction is encoded as a large |
---|
23 | jump is always possible, and the algorithm only reduces those branch |
---|
24 | instructions whose destination address is in range for a shorter jump. |
---|
25 | The algorithm can thus be stopped after a determined number of steps without |
---|
26 | sacrificing correctness. |
---|
27 | |
---|
28 | The result, however, is not necessarily optimal. Even if the algorithm is run |
---|
29 | until it terminates naturally, the fixed point reached is the {\em greatest} |
---|
30 | fixed point, not the least fixed point. Furthermore, {\tt gcc} (at least for |
---|
31 | the x86 architecture) only uses short and long jumps. This makes the algorithm |
---|
32 | more efficient, as shown in the previous section, but also results in a less |
---|
33 | optimal solution. |
---|
34 | |
---|
35 | In the CerCo assembler, we opted at first for a least fixed point algorithm, |
---|
36 | taking absolute jumps into account. |
---|
37 | |
---|
38 | Here, we ran into a problem with proving termination, as explained in the |
---|
39 | previous section: if we only take short and long jumps into account, the jump |
---|
40 | encoding can only switch from short to long, but never in the other direction. |
---|
41 | When we add absolute jumps, however, it is theoretically possible for a branch |
---|
42 | instruction to switch from absolute to long and back, as previously explained. |
---|
43 | Proving termination then becomes difficult, because there is nothing that |
---|
44 | precludes a branch instruction from oscillating back and forth between absolute |
---|
45 | and long jumps indefinitely. |
---|
46 | |
---|
47 | To keep the algorithm in the same complexity class and more easily |
---|
48 | prove termination, we decided to explicitly enforce the `branch instructions |
---|
49 | must always grow longer' requirement: if a branch instruction is encoded as a |
---|
50 | long jump in one iteration, it will also be encoded as a long jump in all the |
---|
51 | following iterations. Therefore the encoding of any branch instruction |
---|
52 | can change at most two times: once from short to absolute (or long), and once |
---|
53 | from absolute to long. |
---|
54 | |
---|
55 | There is one complicating factor. Suppose that a branch instruction is encoded |
---|
56 | in step $n$ as an absolute jump, but in step $n+1$ it is determined that |
---|
57 | (because of changes elsewhere) it can now be encoded as a short jump. Due to |
---|
58 | the requirement that the branch instructions must always grow longer, |
---|
59 | the branch encoding will be encoded as an absolute jump in step |
---|
60 | $n+1$ as well. |
---|
61 | |
---|
62 | This is not necessarily correct. A branch instruction that can be |
---|
63 | encoded as a short jump cannot always also be encoded as an absolute jump, as a |
---|
64 | short jump can bridge segments, whereas an absolute jump cannot. Therefore, |
---|
65 | in this situation we have decided to encode the branch instruction as a long |
---|
66 | jump, which is always correct. |
---|
67 | |
---|
68 | The resulting algorithm, therefore, will not return the least fixed point, as |
---|
69 | it might have too many long jumps. However, it is still better than the |
---|
70 | algorithms from SDCC and {\tt gcc}, since even in the worst case, it will still |
---|
71 | return a smaller or equal solution. |
---|
72 | |
---|
73 | Experimenting with our algorithm on the test suite of C programs included with |
---|
74 | gcc 2.3.3 has shown that on average, about 25 percent of jumps are encoded as short or absolute jumps by the algorithm. As not all instructions are jumps, this does not make for a large reduction in size, but it can make for a reduction in execution time: if jumps |
---|
75 | are executed multiple times, for example in loops, the fact that short jumps take less cycles to execute than long jumps can have great effect. |
---|
76 | |
---|
77 | As for complexity, there are at most $2n$ iterations, where $n$ is the number of |
---|
78 | branch instructions. Practical tests within the CerCo project on small to |
---|
79 | medium pieces of code have shown that in almost all cases, a fixed point is |
---|
80 | reached in 3 passes. Only in one case did the algorithm need 4. This is not surprising: after all, the difference between short/absolute and |
---|
81 | long jumps is only one byte (three for conditional jumps). For a change from |
---|
82 | short/absolute to long to have an effect on other jumps is therefore relatively |
---|
83 | uncommon, which explains why a fixed point is reached so quickly. |
---|
84 | |
---|
85 | \subsection{The algorithm in detail} |
---|
86 | |
---|
87 | The branch displacement algorithm forms part of the translation from |
---|
88 | pseudocode to assembler. More specifically, it is used by the function that |
---|
89 | translates pseudo-addresses (natural numbers indicating the position of the |
---|
90 | instruction in the program) to actual addresses in memory. Note that in pseudocode, all instructions are of size 1. |
---|
91 | |
---|
92 | Our original intention was to have two different functions, one function |
---|
93 | $\mathtt{policy}: \mathbb{N} \rightarrow \{\mathtt{short\_jump}, |
---|
94 | \mathtt{absolute\_jump}, \mathtt{long\_jump}\}$ to associate jumps to their |
---|
95 | intended encoding, and a function $\sigma: \mathbb{N} \rightarrow |
---|
96 | \mathtt{Word}$ to associate pseudo-addresses to machine addresses. $\sigma$ |
---|
97 | would use $\mathtt{policy}$ to determine the size of jump instructions. This turned out to be suboptimal from the algorithmic point of view and |
---|
98 | impossible to prove correct. |
---|
99 | |
---|
100 | From the algorithmic point of view, in order to create the $\mathtt{policy}$ |
---|
101 | function, we must necessarily have a translation from pseudo-addresses |
---|
102 | to machine addresses (i.e. a $\sigma$ function): in order to judge the distance |
---|
103 | between a jump and its destination, we must know their memory locations. |
---|
104 | Conversely, in order to create the $\sigma$ function, we need to have the |
---|
105 | $\mathtt{policy}$ function, otherwise we do not know the sizes of the jump |
---|
106 | instructions in the program. |
---|
107 | |
---|
108 | Much the same problem appears when we try to prove the algorithm correct: the |
---|
109 | correctness of $\mathtt{policy}$ depends on the correctness of $\sigma$, and |
---|
110 | the correctness of $\sigma$ depends on the correctness of $\mathtt{policy}$. |
---|
111 | |
---|
112 | We solved this problem by integrating the $\mathtt{policy}$ and $\sigma$ |
---|
113 | algorithms. We now have a function |
---|
114 | $\sigma: \mathbb{N} \rightarrow \mathtt{Word} \times \mathtt{bool}$ which |
---|
115 | associates a pseudo-address to a machine address. The boolean denotes a forced |
---|
116 | long jump; as noted in the previous section, if during the fixed point |
---|
117 | computation an absolute jump changes to be potentially re-encoded as a short |
---|
118 | jump, the result is actually a long jump. It might therefore be the case that |
---|
119 | jumps are encoded as long jumps without this actually being necessary, and this |
---|
120 | information needs to be passed to the code generating function. |
---|
121 | |
---|
122 | The assembler function encodes the jumps by checking the distance between |
---|
123 | source and destination according to $\sigma$, so it could select an absolute |
---|
124 | jump in a situation where there should be a long jump. The boolean is there |
---|
125 | to prevent this from happening by indicating the locations where a long jump |
---|
126 | should be encoded, even if a shorter jump is possible. This has no effect on |
---|
127 | correctness, since a long jump is applicable in any situation. |
---|
128 | |
---|
129 | \begin{figure}[t] |
---|
130 | \small |
---|
131 | \begin{algorithmic} |
---|
132 | \Function{f}{$labels$,$old\_sigma$,$instr$,$ppc$,$acc$} |
---|
133 | \State $\langle added, pc, sigma \rangle \gets acc$ |
---|
134 | \If {$instr$ is a backward jump to $j$} |
---|
135 | \State $length \gets \mathrm{jump\_size}(pc,sigma_1(labels(j)))$ |
---|
136 | \Comment compute jump distance |
---|
137 | \ElsIf {$instr$ is a forward jump to $j$} |
---|
138 | \State $length \gets \mathrm{jump\_size}(pc,old\_sigma_1(labels(j))+added)$ |
---|
139 | \EndIf |
---|
140 | \State $old\_length \gets \mathrm{old\_sigma_1}(ppc)$ |
---|
141 | \State $new\_length \gets \mathrm{max}(old\_length, length)$ |
---|
142 | \Comment length must never decrease |
---|
143 | \State $old\_size \gets \mathrm{old\_sigma_2}(ppc)$ |
---|
144 | \State $new\_size \gets \mathrm{instruction\_size}(instr,new\_length)$ |
---|
145 | \Comment compute size in bytes |
---|
146 | \State $new\_added \gets added+(new\_size-old\_size)$ |
---|
147 | \Comment keep track of total added bytes |
---|
148 | \State $new\_sigma \gets old\_sigma$ |
---|
149 | \State $new\_sigma_1(ppc+1) \gets pc+new\_size$ |
---|
150 | \State $new\_sigma_2(ppc) \gets new\_length$ |
---|
151 | \Comment update $\sigma$ \\ |
---|
152 | \Return $\langle new\_added, pc+new\_size, new\_sigma \rangle$ |
---|
153 | \EndFunction |
---|
154 | \end{algorithmic} |
---|
155 | \caption{The heart of the algorithm} |
---|
156 | \label{f:jump_expansion_step} |
---|
157 | \end{figure} |
---|
158 | |
---|
159 | The algorithm, shown in Figure~\ref{f:jump_expansion_step}, works by folding the |
---|
160 | function {\sc f} over the entire program, thus gradually constructing $sigma$. |
---|
161 | This constitutes one step in the fixed point calculation; successive steps |
---|
162 | repeat the fold until a fixed point is reached. We have abstracted away the case where an instruction is not a jump, since the size of these instructions is constant. |
---|
163 | |
---|
164 | Parameters of the function {\sc f} are: |
---|
165 | \begin{itemize} |
---|
166 | \item a function $labels$ that associates a label to its pseudo-address; |
---|
167 | \item $old\_sigma$, the $\sigma$ function returned by the previous |
---|
168 | iteration of the fixed point calculation; |
---|
169 | \item $instr$, the instruction currently under consideration; |
---|
170 | \item $ppc$, the pseudo-address of $instr$; |
---|
171 | \item $acc$, the fold accumulator, which contains $added$ (the number of |
---|
172 | bytes added to the program size with respect to the previous iteration), $pc$ |
---|
173 | (the highest memory address reached so far), and of course $sigma$, the |
---|
174 | $\sigma$ function under construction. |
---|
175 | \end{itemize} |
---|
176 | The first two are parameters that remain the same through one iteration, the |
---|
177 | final three are standard parameters for a fold function (including $ppc$, |
---|
178 | which is simply the number of instructions of the program already processed). |
---|
179 | |
---|
180 | The $\sigma$ functions used by {\sc f} are not of the same type as the final |
---|
181 | $\sigma$ function: they are of type |
---|
182 | $\sigma: \mathbb{N} \rightarrow \mathbb{N} \times \{\mathtt{short\_jump}, |
---|
183 | \mathtt{absolute\_jump},\mathtt{long\_jump}\}$; a function that associates a |
---|
184 | pseudo-address with a memory address and a jump length. We do this to |
---|
185 | ease the comparison of jump lengths between iterations. In the algorithm, |
---|
186 | we use the notation $sigma_1(x)$ to denote the memory address corresponding to |
---|
187 | $x$, and $sigma_2(x)$ for the jump length corresponding to $x$. |
---|
188 | |
---|
189 | Note that the $\sigma$ function used for label lookup varies depending on |
---|
190 | whether the label is behind our current position or ahead of it. For |
---|
191 | backward branches, where the label is behind our current position, we can use |
---|
192 | $sigma$ for lookup, since its memory address is already known. However, for |
---|
193 | forward branches, the memory address of the address of the label is not yet |
---|
194 | known, so we must use $old\_sigma$. |
---|
195 | |
---|
196 | We cannot use $old\_sigma$ without change: it might be the case that we have |
---|
197 | already increased the size of some branch instructions before, making the |
---|
198 | program longer and moving every instruction forward. We must compensate for this |
---|
199 | by adding the size increase of the program to the label's memory address |
---|
200 | according to $old\_sigma$, so that branch instruction spans do not get |
---|
201 | compromised. |
---|
202 | |
---|
203 | %Note also that we add the pc to $sigma$ at location $ppc+1$, whereas we add the |
---|
204 | %jump length at location $ppc$. We do this so that $sigma(ppc)$ will always |
---|
205 | %return a pair with the start address of the instruction at $ppc$ and the |
---|
206 | %length of its branch instruction (if any); the end address of the program can |
---|
207 | %be found at $sigma(n+1)$, where $n$ is the number of instructions in the |
---|
208 | %program. |
---|