Hi!
The following testcase ICEs starting with my change to move offsetof-like
expression handling from the parsing to cp_fold - if the base expression
is not INTEGER_CST, but a constant VAR_DECL initialized with INTEGER_CST,
then we don't fold it as offsetof-like expression and as we use recursive
cp_fold only on functions, not initializers, we don't fold that VAR_DECL in
there into INTEGER_CST and the middle-end ICEs on it trying to fold it.
Fixed thusly, bootstrapped/regtested on x86_64-linux and i686-linux, ok for
trunk?
2018-12-09 Jakub Jelinek <ja...@redhat.com>
PR c++/88410
* cp-gimplify.c (cp_fold) <case ADDR_EXPR>: For offsetof-like folding,
call maybe_constant_value on val to see if it is INTEGER_CST.
* g++.dg/cpp0x/pr88410.C: New test.
--- gcc/cp/cp-gimplify.c.jj 2018-11-23 20:00:26.603273556 +0100
+++ gcc/cp/cp-gimplify.c 2018-12-08 11:39:03.983985326 +0100
@@ -2317,6 +2317,7 @@ cp_fold (tree x)
{
val = TREE_OPERAND (val, 0);
STRIP_NOPS (val);
+ val = maybe_constant_value (val);
if (TREE_CODE (val) == INTEGER_CST)
return fold_offsetof (op0, TREE_TYPE (x));
}
--- gcc/testsuite/g++.dg/cpp0x/pr88410.C.jj 2018-12-08 11:41:10.946927148
+0100
+++ gcc/testsuite/g++.dg/cpp0x/pr88410.C 2018-12-08 11:40:48.476291414
+0100
@@ -0,0 +1,7 @@
+// PR c++/88410
+// { dg-do compile { target c++11 } }
+
+typedef __UINTPTR_TYPE__ uintptr_t;
+const uintptr_t a = 32;
+struct C { int b; int c; };
+uintptr_t d { uintptr_t (&reinterpret_cast<C *>(a)->c) - a };
Jakub
