This is a regression present on the mainline and 8 branch: in some specific
circumstances, you can get a crash when max_size is called from the gimplifier
on very dynamic record types. This makes the function a bit more robust.
Tested on x86_64-suse-linux, applied on mainline and 8 branch.
2019-02-08 Eric Botcazou <ebotca...@adacore.com>
* gcc-interface/utils.c (max_size) <tcc_unary>: Be prepared for an
operand with VOID_TYPE.
--
Eric Botcazou
Index: gcc-interface/utils.c
===================================================================
--- gcc-interface/utils.c (revision 268508)
+++ gcc-interface/utils.c (working copy)
@@ -3642,7 +3642,10 @@ fntype_same_flags_p (const_tree t, tree
/* EXP is an expression for the size of an object. If this size contains
discriminant references, replace them with the maximum (if MAX_P) or
- minimum (if !MAX_P) possible value of the discriminant. */
+ minimum (if !MAX_P) possible value of the discriminant.
+
+ Note that the expression may have already been gimplified,in which case
+ COND_EXPRs have VOID_TYPE and no operands, and this must be handled. */
tree
max_size (tree exp, bool max_p)
@@ -3714,11 +3717,15 @@ max_size (tree exp, bool max_p)
return build_int_cst (type, max_p ? 1 : 0);
case tcc_unary:
+ op0 = TREE_OPERAND (exp, 0);
+
if (code == NON_LVALUE_EXPR)
- return max_size (TREE_OPERAND (exp, 0), max_p);
+ return max_size (op0, max_p);
+
+ if (VOID_TYPE_P (TREE_TYPE (op0)))
+ return max_p ? TYPE_MAX_VALUE (type) : TYPE_MIN_VALUE (type);
- op0 = max_size (TREE_OPERAND (exp, 0),
- code == NEGATE_EXPR ? !max_p : max_p);
+ op0 = max_size (op0, code == NEGATE_EXPR ? !max_p : max_p);
if (op0 == TREE_OPERAND (exp, 0))
return exp;