[Bug tree-optimization/79054] missing range information with INT_MAX

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=79054

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
I get

> ./cc1 -quiet t.c -O2 -Wall -Walloc-size-larger-than=1234 -fdump-tree-all-alias
t.c: In function ‘foo’:
t.c:18:3: warning: argument 1 range [1236, 2147483647] exceeds maximum object
size 1234 [-Walloc-size-larger-than=]
   sink (alloc (anti_range (INT_MIN + 2, 1235)));
   ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
t.c:5:7: note: in a call to allocation function ‘alloc’ declared here
 void* alloc (int) __attribute__ ((alloc_size (1)));
       ^~~~~
t.c: In function ‘bar’:
t.c:23:3: warning: argument 1 range [-2147483648, -1] is negative
[-Walloc-size-larger-than=]
   sink (alloc (anti_range (0, INT_MAX)));
   ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
t.c:5:7: note: in a call to allocation function ‘alloc’ declared here
 void* alloc (int) __attribute__ ((alloc_size (1)));
       ^~~~~
t.c: In function ‘foobar’:
t.c:28:3: warning: argument 1 range [1236, 2147483647] exceeds maximum object
size 1234 [-Walloc-size-larger-than=]
   sink (alloc (anti_range (INT_MIN + 2, 1235)));
   ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
t.c:5:7: note: in a call to allocation function ‘alloc’ declared here
 void* alloc (int) __attribute__ ((alloc_size (1)));
       ^~~~~
t.c:29:3: warning: argument 1 range [-2147483648, -1] is negative
[-Walloc-size-larger-than=]
   sink (alloc (anti_range (0, INT_MAX)));
   ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
t.c:5:7: note: in a call to allocation function ‘alloc’ declared here
 void* alloc (int) __attribute__ ((alloc_size (1)));
       ^~~~~

with r244305, so not confirmed (x86_64-*-*).

So VRP does

  <bb 2> [100.00%]:
  # USE = nonlocal
  # CLB = nonlocal
  val_10 = value ();
  # RANGE [0, 1]
  _11 = val_10 >= -2147483646;
  # RANGE [0, 1]
  _12 = val_10 <= 1235;
  # RANGE [0, 1]
  _13 = _11 & _12;
  if (_13 != 0)
    goto <bb 3>; [50.00%]
  else
    goto <bb 4>; [50.00%]

  <bb 3> [50.00%]:

  <bb 4> [100.00%]:
  # val_14 = PHI <val_10(2), -2147483647(3)>

gets asserts like

  <bb 2> [100.00%]:
  # USE = nonlocal
  # CLB = nonlocal
  val_10 = value ();
  # RANGE [0, 1]
  _11 = val_10 >= -2147483646;
  # RANGE [0, 1]
  _12 = val_10 <= 1235;
  # RANGE [0, 1]
  _13 = _11 & _12;
  if (_13 != 0)
    goto <bb 3>; [50.00%]
  else
    goto <bb 4>; [50.00%]

  <bb 3> [50.00%]:
  val_19 = ASSERT_EXPR <val_10, val_10 >= -2147483646>;
  val_20 = ASSERT_EXPR <val_19, val_19 <= 1235>;

  <bb 4> [100.00%]:
  # val_14 = PHI <val_10(2), -2147483647(3)>

so you can see that for val_14 we do not end up adding asserts for val_10
(because we do not rewrite PHIs).  That's likely becuase we are missing
to handle the "false" edge from _13 != 0 with the & handling.  OTOH
val_10 >= -214783646 & val_10 <= 1235 could have been folded to the
simpler (unsigned)val_10 + CST >= CST.

That is, fold-const.c's range_binop code needs (partial) moving to match.pd.