@@ -4140,13 +4140,13 @@ cxx_eval_switch_expr (const constexpr_ctx *ctx, tree t,
bool *non_constant_p, bool *overflow_p,
tree *jump_target)
{
- tree cond = TREE_OPERAND (t, 0);
+ tree cond = SWITCH_COND (t);
cond = cxx_eval_constant_expression (ctx, cond, false,
non_constant_p, overflow_p);
VERIFY_CONSTANT (cond);
*jump_target = cond;
- tree body = TREE_OPERAND (t, 1);
+ tree body = SWITCH_BODY (t);
constexpr_ctx new_ctx = *ctx;
constexpr_switch_state css = css_default_not_seen;
new_ctx.css_state = &css;
@@ -4681,6 +4681,7 @@ cxx_eval_constant_expression (const constexpr_ctx *ctx, tree t,
case COND_EXPR:
if (jump_target && *jump_target)
{
+ tree orig_jump = *jump_target;
/* When jumping to a label, the label might be either in the
then or else blocks, so process then block first in skipping
mode first, and if we are still in the skipping mode at its end,
@@ -4688,7 +4689,19 @@ cxx_eval_constant_expression (const constexpr_ctx *ctx, tree t,
r = cxx_eval_constant_expression (ctx, TREE_OPERAND (t, 1),
lval, non_constant_p, overflow_p,
jump_target);
- if (*jump_target)
+ /* It's possible that we found the label in the then block. But
+ it could have been followed by another jumping statement, e.g.
+ say we're looking for case 1:
+ if (cond)
+ {
+ // skipped statements
+ case 1:; // clears up *jump_target
+ return 1; // and sets it to a RETURN_EXPR
+ }
+ else { ... }
+ in which case we need not go looking to the else block.
+ (goto is not allowed in a constexpr function.) */
+ if (*jump_target && *jump_target == orig_jump)
r = cxx_eval_constant_expression (ctx, TREE_OPERAND (t, 2),
lval, non_constant_p, overflow_p,
jump_target);
new file mode 100644
@@ -0,0 +1,71 @@
+// PR c++/88983
+// { dg-do compile { target c++14 } }
+
+constexpr int
+fn1 (int ay)
+{
+ switch (ay)
+ {
+ if (1)
+ {
+ case 1:
+ return 1;
+ }
+ else
+ {
+ default:;
+ }
+ }
+
+ return 0;
+}
+
+constexpr int
+fn2 (int ay)
+{
+ switch (ay)
+ {
+ if (1)
+ {
+ case 1:
+ break;
+ }
+ else
+ {
+ default:;
+ }
+ }
+
+ return 0;
+}
+
+constexpr int
+fn3 (int ay)
+{
+ int i = 0;
+ while (i++ < 100)
+ {
+ if (i == 1)
+ return 1;
+ switch (ay)
+ {
+ if (1)
+ {
+ case 1:
+ continue;
+ }
+ else
+ {
+ default:;
+ return -1;
+ }
+ }
+ return -1;
+ }
+
+ return -1;
+}
+
+static_assert (fn1 (1) == 1, "");
+static_assert (fn2 (1) == 0, "");
+static_assert (fn3 (1) == 1, "");
Here we have switch (1) with body if (1) { case 1:; return <retval> = 1; } else { default:; } so we're looking for "case 1:", which we found in the then branch while evaluating the COND_EXPR, so *jump_target is cleared. But it's followed by a jumping statement so *jump_target is set to a RETURN_EXPR, which as the new comment describes, confuses the code in <case COND_EXPR>. label_matches then crashes if it gets a RETURN_EXPR as a jump_target. I've expanded the original testcase so as to cover break/continue after the label. I imagine this could also be fixed differently -- don't call label_matches in cxx_eval_constant_expression if jump_target isn't a label/switch case. Bootstrapped/regtested on x86_64-linux, ok for trunk? I'd like to commit to 8 too. 2019-01-31 Marek Polacek <polacek@redhat.com> PR c++/88983 - ICE with switch in constexpr function. * constexpr.c (cxx_eval_switch_expr): Use SWITCH_COND and SWITCH_BODY. (cxx_eval_constant_expression) <case COND_EXPR>: Don't look for the label in the else branch if we found it in the then branch. * g++.dg/cpp1y/constexpr-88983.C: New test.