[Bug tree-optimization/92868] [10 Regression] ICE: tree check: expected integer_cst, have ssa_name in to_wide, at tree.h:5855
jakub at gcc dot gnu.org
gcc-bugzilla@gcc.gnu.org
Mon Dec 9 15:20:00 GMT 2019
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92868
Jakub Jelinek <jakub at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
CC| |jakub at gcc dot gnu.org
--- Comment #2 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
That code has lots of issues. Some in patch form as I was going through them:
gimple_call_alloc_size: for the case when size is INTEGER_CST and it is the one
argument alloc_size, I don't see a reason not to return size earlier, there is
no need to wi::to_wide. I don't see a point in computing rng1[0] and rng2[0]
when nothing really uses it (write-only element). Some formatting and typo in
comment.
I don't understand why you are using wi::sign_mask, the normal test for
negative wide_int when it is treated as signed integer of the corresponding
precision is wi::neg_p, that is more readable and descriptive. If something is
tested on the wide ints with extra precision and the ::from was UNSIGNED, that
is of course not possible, but sign_mask will not work in that case either.
And if it is ::from with SIGNED, neg_p should work fine too.
There is a weird:
wide_int declsize = wi::to_wide (size);
if (wi::sign_mask (dstoffrng[0]) > 0)
declsize += dstoffrng[0];
that condition is never true, sign_mask will only ever return 0 or -1:
return (HOST_WIDE_INT) (high) < 0 ? -1 : 0;
None of this fixes the ICE. How exactly to fix that depends on whether *poff
can be something other than INTEGER_CST or not. If it can be only INTEGER_CST,
then whatever code is setting *poff to non-INTEGER_CST should instead punt or
set to some safe value, whatever, if it can be anything, then while it is fine
to call functions like integer_zerop etc. on that, tree_int_cst_sgn requires
the argument to be INTEGER_CST only, so there needs to be TREE_CODE (*poff) ==
INTEGER_CST && tree_int_csgn (*poff) < 0 instead. Or perhaps you want
!tree_expr_nonnegative_p (*poff) instead?
--- gcc/builtins.c.jj 2019-12-05 09:47:23.178710510 +0100
+++ gcc/builtins.c 2019-12-09 15:52:55.951404452 +0100
@@ -3746,36 +3746,33 @@ gimple_call_alloc_size (gimple *stmt)
}
tree size = gimple_call_arg (stmt, argidx1);
+ if (argidx2 > nargs && TREE_CODE (size) == INTEGER_CST)
+ return size;
wide_int rng1[2];
if (TREE_CODE (size) == INTEGER_CST)
- rng1[0] = rng1[1] = wi::to_wide (size);
+ rng1[1] = wi::to_wide (size);
else if (TREE_CODE (size) != SSA_NAME
|| get_range_info (size, rng1, rng1 + 1) != VR_RANGE)
return NULL_TREE;
- if (argidx2 > nargs && TREE_CODE (size) == INTEGER_CST)
- return size;
-
/* To handle ranges do the math in wide_int and return the product
of the upper bounds as a constant. Ignore anti-ranges. */
- tree n = argidx2 < nargs ? gimple_call_arg (stmt, argidx2) :
integer_one_node;
+ tree n
+ = argidx2 < nargs ? gimple_call_arg (stmt, argidx2) : integer_one_node;
wide_int rng2[2];
if (TREE_CODE (n) == INTEGER_CST)
- rng2[0] = rng2[1] = wi::to_wide (n);
+ rng2[1] = wi::to_wide (n);
else if (TREE_CODE (n) != SSA_NAME
|| get_range_info (n, rng2, rng2 + 1) != VR_RANGE)
return NULL_TREE;
- /* Extend to the maximum precsion to avoid overflow. */
+ /* Extend to the maximum precision to avoid overflow. */
const int prec = ADDR_MAX_PRECISION;
- rng1[0] = wide_int::from (rng1[0], prec, UNSIGNED);
rng1[1] = wide_int::from (rng1[1], prec, UNSIGNED);
- rng2[0] = wide_int::from (rng2[0], prec, UNSIGNED);
rng2[1] = wide_int::from (rng2[1], prec, UNSIGNED);
/* Return the lesser of SIZE_MAX and the product of the upper bounds. */
- rng1[0] = rng1[0] * rng2[0];
rng1[1] = rng1[1] * rng2[1];
tree size_max = TYPE_MAX_VALUE (sizetype);
if (wi::gtu_p (rng1[1], wi::to_wide (size_max, prec)))
@@ -3853,7 +3850,7 @@ compute_objsize (tree dest, int ostype,
/* Ignore negative offsets for now. For others,
use the lower bound as the most optimistic
estimate of the (remaining) size. */
- if (wi::sign_mask (wioff))
+ if (wi::neg_p (wioff))
;
else if (wi::ltu_p (wioff, wisiz))
{
@@ -3882,9 +3879,8 @@ compute_objsize (tree dest, int ostype,
/* Ignore negative offsets for now. For others,
use the lower bound as the most optimistic
- estimate of the (remaining)size. */
- if (wi::sign_mask (min)
- || wi::sign_mask (max))
+ estimate of the (remaining) size. */
+ if (wi::neg_p (min) || wi::neg_p (max))
;
else if (wi::ltu_p (min, wisiz))
{
@@ -3912,8 +3908,7 @@ compute_objsize (tree dest, int ostype,
if (!ostype)
return NULL_TREE;
- if (TREE_CODE (dest) == ARRAY_REF
- || TREE_CODE (dest) == MEM_REF)
+ if (TREE_CODE (dest) == ARRAY_REF || TREE_CODE (dest) == MEM_REF)
{
tree ref = TREE_OPERAND (dest, 0);
tree off = TREE_OPERAND (dest, 1);
@@ -3972,7 +3967,7 @@ compute_objsize (tree dest, int ostype,
*poff = size_binop (PLUS_EXPR, *poff, off);
}
- if (wi::sign_mask (offrng[0]) >= 0)
+ if (!wi::neg_p (offrng[0]))
{
if (TREE_CODE (size) != INTEGER_CST)
return NULL_TREE;
@@ -4007,7 +4002,7 @@ compute_objsize (tree dest, int ostype,
declsize += dstoffrng[0];
offrng[1] += dstoffrng[1];
- if (wi::sign_mask (offrng[1]) < 0)
+ if (wi::neg_p (offrng[1]))
return size_zero_node;
return wide_int_to_tree (sizetype, declsize);
More information about the Gcc-bugs
mailing list