Logika: Programming Logics
1. Introduction: Why bother?

1. Introduction: Why bother?

That’s a good question! After all, you can hack a spreadsheet program or build an interactive game by writing a lot of code, experimenting with it, and patching it. After awhile, the program you wrote does more or less what you wanted.

But imagine if the rest of the world worked that same way – would you want to drive a car or fly an airplane that was “hacked together”? How about travelling in a bus across a bridge that fell down a few times already and was repeatedly patched till it (seemed) to hold?

Perhaps these analogies are a bit extreme, but professional scientists and engineers rely on planning, design, and calculation so that they are certain the products they want to build will work before anyone starts building them. Professionals rely on an intellectual foundation to plan, design, and calculate. For example:

  • physicists use mathematics as the intellectual foundation of their products

  • chemical engineers use chemistry

  • mechanical engineers use physics

  • computer engineers and computer (software) scientists use algebra and symbolic logic

If you develop significant expertise in software engineering, perhaps you will work at a firm or lab that develops safety-critical software, that is, software upon which people’s money or safety or lives depend. (An example is the flight-control software that lives in a jet. Another example is the navigation software in a satellite that talks to the GPS device in someone’s car.) Software of this nature has to be working correctly from the beginning – there is no freedom to hack-and-patch the code once it is in use. Software engineers must use algebra and logic to plan and calculate how the software will behave before the software is built and installed.

This story is not an idle one: As you probably know, computer processor chips are planned out in a programming language that looks a lot like C. When Intel designed its first Pentium chip, there was a programming error in one of the chip’s coded hash tables. The coding was burned into hardware, and millions of chips were manufactured. The error was quickly detected – the chip did not always perform multiplication correctly. As a result, Intel lost a lot of money recalling the faulty chips and manufacturing a patched replacement. These days, Intel uses techniques for validating chip designs much like the ones you will learn in this course.

If you have taken a software architecture course (e.g., CIS 501: Software Architecture and Design), you know that large systems can be drawn out, or “blueprinted,” with diagrams that show the components and how they connect together by means of method calls, event broadcast, and message passing. What we will learn in this course is lower level and more basic – we will learn how to calculate how the lines of coding in each component compute internal knowledge as they convert inputs into outputs.

To understand the idea, let’s think about electronics. When an electronic device, like a TV-set or radio, is designed, the parts of the device and their wirings are drawn out in a diagram called a schematic. Here is a schematic of a vacuum-tube guitar amplifier, the kind used by recording studios to produce a warm sound with good sustain:

../../_images/eicohf22.png

Notice that the wires to the vacuum tubes (the globes labelled V1 through V5) are labelled with voltages, and there is a table in the lower left corner of the schematic that lists the correct resistances that will hold at each of the wires (“pins”) that connect to the tubes.

The voltage and resistance calculations are both an analysis and a prediction of how the circuit should behave. The numbers were calculated with mathematics and algebra, and if the electronics parts are working correctly, then these voltage, amperage, and resistance levels must occur – the foundations of electronics (math and algebra) demand it.

When the circuit is built, the actual levels are measured with a multimeter and compared to the calculations; if there is a discrepency, this is a signal that some part within the circuit is faulty.

A computer program is a “circuit” that “runs on” knowledge, and when we design the parts (lines) of a computer program, we should include “knowledge checks” that assert the amount of knowledge computed by the program at various points. We will learn how to write and insert such knowledge checks, called assertions, into programs and use the laws of symbolic logic to prove that the assertions will hold true.

You will see many examples of “program schematics” in the upcoming chapters. Here are two. First, this little code fragment apparently selects the larger of two integers and prints it:

 1import org.sireum.logika._
 2// the above imports, for example, type Z,
 3// which is an arbitrary-precision integer type (i.e., scala.BigInt)
 4
 5val x: Z = readInt() // readInt asks an integer from the user via console input
 6val y: Z = readInt() // val declares a read-only variable
 7var max: Z = 0       // var declares a read/write variable
 8if (x > y) {
 9  max = x
10} else {
11  max = y
12}
13println("Maximum of ", x, " and ", y, " is ", max, ".")

Think of the program as a “circuit” whose lines are “wired” together in sequence. Instead of voltage, information or knowledge “flows” from one line to the next. Here is the program’s “schematic” where the internal “knowledge levels” are written in symbolic logic and are inserted within the lines of the program, enclosed by set braces, l"""{ ... }""":

 3val x: Z = readInt()
 4val y: Z = readInt()
 5var max: Z = 0
 6if (x > y) {
 7  l"""{ 1. x > y                   premise     }"""
 8  max = x
 9  l"""{ 1. x > y                   premise
10        2. max == x                premise
11        3. max ≥ x                 algebra 2
12        4. max ≥ y                 algebra 1 3
13        5. max ≥ x  ∧  max ≥ y     ∧i 3 4      }"""
14} else {
15  l"""{ 1. ¬(x > y)                premise
16        2. y ≥ x                   algebra 1   }"""
17  max = y
18  l"""{ 1. max == y                premise
19        2. y ≥ x                   premise
20        3. max ≥ y                 algebra 1
21        4. max ≥ x                 algebra 1 2
22        5. max ≥ x  ∧  max ≥ y     ∧i 4 3      }"""
23}
24l"""{ 1. max ≥ x  ∧  max ≥ y       premise     }"""
25println("Maximum of ", x, " and ", y, " is ", max, ".")

The last annotation, l"""{ ... max >= x max >= y ... }""", is a symbolic-logic statement that max is guaranteed to be greater-or-equal to both inputs. We now know, once the program is implemented, it will behave with this logical property.

Here is a second example, a complete analysis of a function that squares all the integers in an array that is passed to it as its argument:

  3// Updates parameter a, which is of type array of integers (ZS),
  4// in place so that each of its ints are squared
  5def square(a: ZS): Unit = {
  6  l"""{ modifies     a
  7        post         ∀i: (0 ..< a.size)  a(i) == a_in(i) * a_in(i)              }"""
  8
  9  l"""{ 1. a == a_in                                         premise            }"""
 10
 11  var x: Z = 0
 12
 13  l"""{ 1. a == a_in                                         premise
 14        2. x == 0                                            premise
 15        3. 0 ≤ x                                             algebra 2
 16        4. 0 ≤ a.size                                        algebra
 17        5. x ≤ a.size                                        subst2 2 4
 18        6. {
 19             7. j: Z
 20             8. {
 21                  9. 0 ≤ j ∧ j < x                           assume
 22                 10. 0 ≤ j                                   ∧e1 9
 23                 11. j < x                                   ∧e2 9
 24                 12. ⊥                                       algebra 10 11 2
 25                 13. a(j) == a_in(j) * a_in(j)               ⊥e 12
 26                }
 27            14. 0 ≤ j ∧ j < x → a(j) == a_in(j) * a_in(j)    →i 8
 28           }
 29       15. ∀i: (0 ..< x)  a(i) == a_in(i) * a_in(i)          ∀i 6
 30       16. {
 31            17. j: Z
 32            18. {
 33                 19. x ≤ j ∧ j < a.size                      assume
 34                 20. x ≤ j                                   ∧e1 19
 35                 21. j < a.size                              ∧e2 19
 36                 22. a(j) == a_in(j)                         algebra 1 2 20 21
 37                }
 38            23. x ≤ j ∧ j < a.size → a(j) == a_in(j)         →i 18
 39           }
 40       24. ∀i: (x ..< a.size)  a(i) == a_in(i)               ∀i 16              }"""
 41
 42  while (x != a.size) {
 43    l"""{ invariant  ∀i: (0 ..< x)  a(i) == a_in(i) * a_in(i)
 44                     ∀i: (x ..< a.size)  a(i) == a_in(i)
 45                     0 ≤ x
 46                     x ≤ a.size
 47          modifies   x, a                                                       }"""
 48
 49    l"""{ 1. ∀i: (0 ..< x)  a(i) == a_in(i) * a_in(i)        premise
 50          2. ∀i: (x ..< a.size)  a(i) == a_in(i)             premise
 51          3. 0 ≤ x                                           premise
 52          4. x ≤ a.size                                      premise
 53          5. x ≠ a.size                                      premise
 54          6. x < a.size                                      algebra 4 5        }"""
 55
 56    a(x) = a(x) * a(x)
 57
 58    l"""{ 1. a(x) == a_old(x) * a_old(x)                     premise
 59          2. ∀i: (0 ..< x)  a_old(i) == a_in(i) * a_in(i)    premise
 60          3. ∀i: (x ..< a_old.size)  a_old(i) == a_in(i)     premise
 61          4. a.size == a_old.size                            premise
 62          5. x < a_old.size                                  premise
 63          6. x < a.size                                      subst2 4 5
 64          7. ∀i: (x ..< a.size)  a_old(i) == a_in(i)         subst2 4 3
 65          8. x ≤ x ∧ x < a.size → a_old(x) == a_in(x)        ∀e 7 x
 66          9. x ≤ x                                           algebra
 67         10. x ≤ x ∧ x < a.size                              ∧i 9 6
 68         11. a_old(x) == a_in(x)                             →e 8 10
 69         12. a(x) == a_in(x) * a_in(x)                       subst1 11 1
 70         13. ∀q_i: (0 ..< a.size)
 71               q_i ≠ x → a(q_i) == a_old(q_i)                premise
 72         14. {
 73              15. j: Z
 74              16. 0 ≤ j ∧ j < a.size →
 75                    (j ≠ x → a(j) == a_old(j))               ∀e 13 j
 76              17. {
 77                   18. 0 ≤ j ∧ j ≤ x                         assume
 78                   19. 0 ≤ j ∧ j < x →
 79                         a_old(j) == a_in(j) * a_in(j)       ∀e 2 j
 80                   20. 0 ≤ j                                 ∧e1 18
 81                   21. j ≤ x                                 ∧e2 18
 82                   22. {
 83                        23. j < x                            assume
 84                        24. 0 ≤ j ∧ j < x                    ∧i 20 23
 85                        25. a_old(j) == a_in(j) * a_in(j)    →e 19 24
 86                        26. j < a.size                       algebra 23 6
 87                        27. j ≠ x                            algebra 23
 88                        29. 0 ≤ j ∧ j < a.size               ∧i 20 26
 89                        30. j ≠ x → a(j) == a_old(j)         →e 16 29
 90                        31. a(j) == a_old(j)                 →e 30 27
 91                        32. a(j) == a_in(j) * a_in(j)        subst2 31 25
 92                       }
 93                   33. {
 94                        34. j == x                           assume
 95                        35. a(j) == a_in(j) * a_in(j)        subst2 34 12
 96                       }
 97                   36. a(j) == a_in(j) * a_in(j)             ∨e 21 22 33
 98                  }
 99              37. 0 ≤ j ∧ j ≤ x → a(j) == a_in(j) * a_in(j)  →i 17
100             }
101         38. ∀i: (0 .. x) a(i) == a_in(i) * a_in(i)          ∀i 14
102         39. 0 ≤ x                                           premise
103         40. {
104              41. j: Z
105              42. 0 ≤ j ∧ j < a.size →
106                    (j ≠ x → a(j) == a_old(j))               ∀e 13 j
107              43. {
108                   44. x + 1 ≤ j ∧ j < a.size                assume
109                   45. x + 1 ≤ j                             ∧e1 44
110                   46. j < a.size                            ∧e2 44
111                   47. 0 ≤ j                                 algebra 45 39
112                   48. j ≠ x                                 algebra 45
113                   49. 0 ≤ j ∧ j < a.size                    ∧i 47 46
114                   50. j ≠ x → a(j) == a_old(j)              →e 42 49
115                   51. a(j) == a_old(j)                      →e 50 48
116                   52. x ≤ j ∧ j < a.size →
117                         a_old(j) == a_in(j)                 ∀e 7 j
118                   53. x ≤ j                                 algebra 45
119                   54. x ≤ j ∧ j < a.size                    ∧i 53 46
120                   55. a_old(j) == a_in(j)                   →e 52 54
121                   56. a(j) == a_in(j)                       subst1 55 51
122                  }
123              57. x + 1 ≤ j ∧ j < a.size → a(j) == a_in(j)   →i 43
124             }
125         58. ∀i: (x + 1 ..< a.size)  a(i) == a_in(i)         ∀i 40              }"""
126
127    x = x + 1
128
129    l"""{ 1. x == x_old + 1                                  premise
130          2. 0 ≤ x_old                                       premise
131          3. x_old < a.size                                  premise
132          4. ∀i: (0 .. x_old)  a(i) == a_in(i) * a_in(i)     premise
133          5. ∀i: (x_old + 1 ..< a.size)  a(i) == a_in(i)     premise
134          6. 0 ≤ x                                           algebra 1 2
135          7. x ≤ a.size                                      algebra 1 3
136          8. ∀i: (x ..< a.size)  a(i) == a_in(i)             subst2 1 5
137          9. {
138              10. j: Z
139              11. 0 ≤ j ∧ j ≤ x_old →
140                    a(j) == a_in(j) * a_in(j)                ∀e 4 j
141              12. {
142                   13. 0 ≤ j ∧ j < x                         assume
143                   14. 0 ≤ j                                 ∧e1 13
144                   15. j < x                                 ∧e2 13
145                   16. j ≤ x_old                             algebra 15 1
146                   17. 0 ≤ j ∧ j ≤ x_old                     ∧i 14 16
147                   18. a(j) == a_in(j) * a_in(j)             →e 11 17
148                  }
149              19. 0 ≤ j ∧ j < x → a(j) == a_in(j) * a_in(j)  →i 12
150             }
151         20. ∀i: (0 ..< x)  a(i) == a_in(i) * a_in(i)        ∀i 9               }"""
152  }
153  l"""{ 1. ∀i: (0 ..< x)  a(i) == a_in(i) * a_in(i)          premise
154        2. not (x != a.size)                                 premise
155        3. x == a.size                                       algebra 2
156        4. ∀i: (0 ..< a.size)  a(i) == a_in(i) * a_in(i)     subst1 3 1         }"""
157}

You are not expected to understand the above, but the function’s post-condition list the guarantees of what goes out for any given array of integers. In this case, “what goes out” is an array whose elements are squared – it is guaranteed to work, because it was analyzed the same way an electronics engineer analyzes a circuit.

If the above somehow looks daunting, here is a shorter proof that leverages Logika’s automation:

 3def square(a: ZS): Unit = {
 4  l"""{ modifies     a
 5        post         ∀i: (0 ..< a.size)  a(i) == a_in(i) * a_in(i)              }"""
 6
 7  var x: Z = 0
 8
 9  while (x != a.size) {
10    l"""{ invariant  ∀i: (0 ..< x)  a(i) == a_in(i) * a_in(i)
11                     ∀i: (x ..< a.size)  a(i) == a_in(i)
12                     0 ≤ x
13                     x ≤ a.size
14          modifies   x, a                                                       }"""
15
16    a(x) = a(x) * a(x)
17
18    l"""{ 1. x < a.size                                                auto
19          2. ∀i: (0 .. x)  a(i) == a_in(i) * a_in(i)                   auto
20          3. ∀i: (x + 1 ..< a.size)  a(i) == a_in(i)                   auto     }"""
21
22    x = x + 1
23  }
24}

Even better, using symbolic execution, it can be automatically proved without proof annotations:

 3def square(a: ZS): Unit = {
 4  l"""{ modifies     a
 5        post         ∀i: (0 ..< a.size)  a(i) == a_in(i) * a_in(i)              }"""
 6
 7  var x: Z = 0
 8
 9  while (x != a.size) {
10    l"""{ invariant  ∀i: (0 ..< x)  a(i) == a_in(i) * a_in(i)
11                     ∀i: (x ..< a.size)  a(i) == a_in(i)
12                     0 ≤ x
13                     x ≤ a.size
14          modifies   x, a                                                       }"""
15
16    a(x) = a(x) * a(x)
17
18    x = x + 1
19  }
20}

This note was adapted from David Schmidt's CIS 301, 2008, Chapter 00 course note.