Add function pointer restriction option

This adds a new option, --restrict-function-pointer, to cbmc. This lets a user
specify a list of possible pointer targets for specific function pointer call
sites, rather than have remove_function_pointers guess possible values. The
intended purpose behind this is to prevent excessive symex time wasted on
exploring paths the user knows the program can never actually take.

Author: hannes-steffenhagen-diffblue

doc/cprover-manual/index.md

@@ -9,6 +9,7 @@
 [A Short Tutorial](cbmc/tutorial/),
 [Loop Unwinding](cbmc/unwinding/),
 [Assertion Checking](cbmc/assertions/),
+[Restricting function pointers](cbmc/restrict-function-pointer/),
 [Memory Analyzer](cbmc/memory-analyzer/),
 [Program Harness](cbmc/goto-harness/)
 

doc/cprover-manual/restrict-function-pointer.md

@@ -0,0 +1,147 @@
+[CPROVER Manual TOC](../../)
+
+## Restricting function pointers
+
+### Motivation
+
+CBMC comes with a way to resolve calls to function pointers to direct function
+calls. This is needed because symbolic execution itself can't handle calls to
+function pointers. In practice, this looks something like this:
+
+Given that there are functions with these signatures available in the program:
+
+```
+int f(int x);
+int g(int x);
+int h(int x);
+```
+
+And we have a call site like this:
+
+```
+typedef int(*fptr_t)(int x);
+void call(fptr_t fptr) {
+  int r = fptr(10);
+  assert(r > 0);
+}
+```
+
+Function pointer removal will turn this into code similar to this:
+
+```
+void call(fptr_t fptr) {
+  int r;
+  if(fptr == &f) {
+    r = f(10);
+  } else if(fptr == &g) {
+    r = g(10);
+  } else if(fptr == &h) {
+    r = h(10);
+  } else {
+    // sanity check
+    assert(false);
+    assume(false);
+  }
+  return r;
+}
+```
+
+This works well enough for simple cases. However, this is a "dumb" replacement
+only based on the signature of the function, so if there are many functions
+matching a particular signature, or if some of these functions are expensive in
+symex (e.g. functions with lots of loops or recursion), then this can be a bit
+cumbersome - especially if we, as a user, already know that a particular
+function pointer will only resolve to a single function or a small set of
+functions. This is what the `--restrict-function-pointer` option allows.
+
+### Example
+
+Take the motivating example. Let us assume that we know for a fact that `call`
+will always receive pointers to either `f` or `g` during actual executions of
+the program, and symbolic execution for `h` is too expensive to simply ignore
+the cost of its branch. For this, we will label the places in each function
+where function pointers are being called, to this pattern:
+
+```
+<function-name>.function_pointer_call.<N>
+```
+
+where `N` is referring to which function call it is - so the first call to a
+function pointer in a function will have `N=1`, the 5th `N=5` etc.
+
+We can call `cbmc` with `--restrict-function-pointer
+call.function_pointer_call.1/f,g`. This can be read as
+
+> For the first call to a function pointer in the function `call`, assume that
+> it can only be a call to `f` or `g`
+
+The resulting code looks similar to the original example, except now there will
+not be a call to `h`:
+
+```
+void call(fptr_t fptr) {
+  int r;
+  if(fptr == &f) {
+    r = f(10);
+  } else if(fptr == &g) {
+    r = g(10);
+  } else {
+    // sanity check
+    assert(false);
+    assume(false);
+  }
+  return r;
+}
+```
+
+Another example: Imagine we have a simple virtual filesystem API and implementation
+like this:
+
+```
+typedef struct filesystem_t filesystem_t;
+struct filesystem_t {
+  int (*open)(filesystem_t *filesystem, const char* file_name);
+};
+
+int fs_open(filesystem_t *filesystem, const char* file_name) {
+  filesystem->open(filesystem, file_name);
+}
+
+int nullfs_open(filesystem_t *filesystem, const char* file_name) {
+  return -1;
+}
+
+filesystem_t nullfs_val = {.open = nullfs_open};
+filesystem *const nullfs = &nullfs_val;
+
+filesystem_t *get_fs_impl() {
+  // some fancy logic to determine
+  // which filesystem we're getting -
+  // in-memory, backed by a database, OS file system
+  // - but in our case, we know that
+  // it always ends up being nullfs
+  // for the cases we care about
+  return nullfs;
+}
+int main(void) {
+  filesystem_t *fs = get_fs_impl();
+  assert(fs_open(fs, "hello.txt") != -1);
+}
+```
+
+In this case, the assumption is that *we* know that in our `main`, `fs` can be
+nothing other than `nullfs`; But perhaps due to the logic being too complicated
+symex ends up being unable to figure this out, so in the call to `fs_open()` we
+end up branching on all functions matching the signature of
+`filesystem_t::open`, which could be quite a few functions within the program.
+Worst of all, if it's address is ever taken in the program, as far as the "dumb"
+function pointer removal is concerned it could be `fs_open()` itself due to it
+having a matching signature, leading to symex being forced to follow a
+potentially infinite recursion until its unwind limit.
+
+In this case we can again restrict the function pointer to the value which we
+know it will have:
+
+```
+--restrict-function-pointer fs_open.function_pointer_call.1/nullfs_open
+```

regression/cbmc/restrict-function-pointer-to-complex-expression/test.c

@@ -0,0 +1,60 @@
+#include <assert.h>
+
+typedef int (*fptr_t)(int);
+
+fptr_t get_f(void);
+fptr_t get_g(void);
+
+void use_fg(int choice, fptr_t fptr, fptr_t gptr)
+{
+  assert((choice ? fptr : gptr)(10) == 10 + choice);
+}
+
+// this is just here to distinguish the behavior from FP removal, which'd include g
+int g_always_false_cond = 0;
+
+int main(void)
+{
+  g_always_false_cond = 1;
+  use_fg(0, get_f(), get_g());
+  use_fg(1, get_f(), get_g());
+}
+
+int f(int x)
+{
+  return x + 1;
+}
+
+int g(int x)
+{
+  return x;
+}
+
+int h(int x)
+{
+  return x / 2;
+}
+
+fptr_t get_f(void)
+{
+  if(!g_always_false_cond)
+  {
+    return f;
+  }
+  else
+  {
+    return h;
+  }
+}
+
+fptr_t get_g(void)
+{
+  if(!g_always_false_cond)
+  {
+    return g;
+  }
+  else
+  {
+    return h;
+  }
+}

regression/cbmc/restrict-function-pointer-to-complex-expression/test.desc

@@ -0,0 +1,66 @@
+#include <assert.h>
+
+typedef int (*fptr_t)(int);
+
+fptr_t get_f(void);
+
+void use_f(fptr_t fptr)
+{
+  assert(fptr(10) >= 10);
+}
+
+void select_f(void);
+void select_g(void);
+void select_h(void);
+
+int main(void)
+{
+  select_f();
+  use_f(get_f());
+  select_g();
+  use_f(get_f());
+}
+
+int f(int x)
+{
+  return x + 1;
+}
+int g(int x)
+{
+  return x;
+}
+int h(int x)
+{
+  return x - 1;
+}
+
+int g_select_function = 0;
+
+void select_f(void)
+{
+  g_select_function = 0;
+}
+void select_g(void)
+{
+  g_select_function = 1;
+}
+void select_h(void)
+{
+  g_select_function = 2;
+}
+
+fptr_t get_f(void)
+{
+  if(g_select_function == 0)
+  {
+    return f;
+  }
+  else if(g_select_function == 1)
+  {
+    return g;
+  }
+  else
+  {
+    return h;
+  }
+}

regression/cbmc/restrict-function-pointer-to-multiple-functions-code-check/test.c

@@ -0,0 +1,9 @@
+CORE
+test.c
+--restrict-function-pointer use_f.function_pointer_call.1/f,g --show-goto-functions
+f\(10\)
+g\(10\)
+^EXIT=0$
+^SIGNAL=0$
+--
+h\(10\)

regression/cbmc/restrict-function-pointer-to-multiple-functions-code-check/test.desc

@@ -0,0 +1,68 @@
+#include <assert.h>
+
+typedef int (*fptr_t)(int);
+
+fptr_t get_f(void);
+
+void use_f(fptr_t fptr)
+{
+  assert(fptr(10) >= 10);
+}
+
+void select_f(void);
+void select_g(void);
+void select_h(void);
+
+int main(void)
+{
+  select_f();
+  use_f(get_f());
+  select_g();
+  use_f(get_f());
+  select_h();
+  use_f(get_f());
+}
+
+int f(int x)
+{
+  return x + 1;
+}
+int g(int x)
+{
+  return x;
+}
+int h(int x)
+{
+  return x - 1;
+}
+
+int g_select_function = 0;
+
+void select_f(void)
+{
+  g_select_function = 0;
+}
+void select_g(void)
+{
+  g_select_function = 1;
+}
+void select_h(void)
+{
+  g_select_function = 2;
+}
+
+fptr_t get_f(void)
+{
+  if(g_select_function == 0)
+  {
+    return f;
+  }
+  else if(g_select_function == 1)
+  {
+    return g;
+  }
+  else
+  {
+    return h;
+  }
+}

regression/cbmc/restrict-function-pointer-to-multiple-functions-incorrectly/test.c

@@ -0,0 +1,6 @@
+CORE
+test.c
+--restrict-function-pointer use_f.function_pointer_call.1/f,g
+\[use_f\.assertion\.1\] line \d+ use_f.function_pointer_call.1 must be one of \[f, g\]: FAILURE
+^EXIT=10$
+^SIGNAL=0$