Hi!
The following testcase ICEs because we get:
(call_insn/u 7 4 14 2 (set (reg:QI 0 ax)
(call (mem:QI (symbol_ref:DI ("_Zlt1AS_") [flags 0x3] <function_decl
0x7efd00f7fc00 operator<>) [0 operator< S1 A8])
(const_int 0 [0]))) 690 {*call_value}
(expr_list:REG_EH_REGION (const_int 0 [0])
(nil))
(expr_list:QI (use (mem:QI (reg/f:DI 7 sp) [0 S1 A64]))
(expr_list:QI (use (mem:QI (reg/f:DI 7 sp) [0 S1 A64]))
(nil))))
where the CALL_INSN_FUNCTION_USAGE USEs for the two empty arguments
have size 1 and overlap. In the PR I've added a patch that makes sure
the MEMs have BLKmode and zero size, but actually I fail to see any
advantages in having anything for the TYPE_EMPTY_P arguments in C_I_F_U,
because we actually don't store anything at all, and furthermore I don't
see even the need to actually create the MEMs for stack and stack_slot
fields etc.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
2017-12-19 Jakub Jelinek <ja...@redhat.com>
PR c++/83490
* calls.c (compute_argument_addresses): Ignore TYPE_EMPTY_P arguments.
* g++.dg/abi/empty29.C: New test.
--- gcc/calls.c.jj 2017-12-18 14:57:24.000000000 +0100
+++ gcc/calls.c 2017-12-19 20:25:00.143659975 +0100
@@ -2365,6 +2365,9 @@ compute_argument_addresses (struct arg_d
&& args[i].partial == 0)
continue;
+ if (TYPE_EMPTY_P (TREE_TYPE (args[i].tree_value)))
+ continue;
+
/* Pointer Bounds are never passed on the stack. */
if (POINTER_BOUNDS_P (args[i].tree_value))
continue;
--- gcc/testsuite/g++.dg/abi/empty29.C.jj 2017-12-19 20:46:48.141253484
+0100
+++ gcc/testsuite/g++.dg/abi/empty29.C 2017-12-19 20:47:05.043041447 +0100
@@ -0,0 +1,7 @@
+// PR c++/83490
+// { dg-do compile { target i?86-*-* x86_64-*-* } }
+// { dg-options "-fipa-icf-functions -Og -maccumulate-outgoing-args" }
+
+struct A {};
+A operator < (A, A) { return A (); }
+A operator > (A, A) { return A (); }
Jakub
