trippels@gcc1-power7 status % cat defaults.ii struct A { }; template <class L, class R> A operator%(L, R); template <class A0, class A1, class A2, class A3> void make_tuple (A0 &, A1, A2, A3); A bar (int p1, char p2, int p3, double p4) { A a; make_tuple (p1, p2, p3, p4); return "int; char; string; double; " % a; } A foo (int p1, char p2, int p3, double p4) { A b; make_tuple (p1, p2, p3, p4); return "int; char; string; double; " % b; } trippels@gcc1-power7 status % g++ -c -w -O2 -fPIC defaults.ii defaults.ii: In function ‘A bar(int, char, int, double)’: defaults.ii:20:1: error: invalid argument to gimple call } ^ p1 # .MEM_5 = VDEF <.MEM_1(D)> <retval> = _Z3fooicid.localalias.0 (p1, p2_2(D), p3_3(D), p4_4(D)); [tail call] defaults.ii:20:1: internal compiler error: verify_gimple failed 0x109d2c2f verify_gimple_in_cfg(function*, bool) ../../gcc/gcc/tree-cfg.c:5025 0x10893493 execute_function_todo ../../gcc/gcc/passes.c:1755 0x10894137 do_per_function ../../gcc/gcc/passes.c:1489 0x10894137 do_per_function ../../gcc/gcc/passes.c:1479 0x108942ef execute_todo ../../gcc/gcc/passes.c:1812 Please submit a full bug report, with preprocessed source if appropriate. Please include the complete backtrace with any bug report. See <http://gcc.gnu.org/bugs.html> for instructions. trippels@gcc1-power7 status % g++ -c -fno-ipa-icf -O2 -fPIC defaults.ii trippels@gcc1-power7 status %
You miss to mark p1 addressable in the alias decl (that is, copy TREE_ADDRESSABLE).
(In reply to Richard Biener from comment #1) > You miss to mark p1 addressable in the alias decl (that is, copy > TREE_ADDRESSABLE). As I can see, ICF creates thunk to a localalias of foo. At which please I should preserve addressable attribute? Thanks, Martin
> https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63580 > > Richard Biener <rguenth at gcc dot gnu.org> changed: > > What |Removed |Added > ---------------------------------------------------------------------------- > Status|UNCONFIRMED |NEW > Last reconfirmed| |2014-10-20 > Ever confirmed|0 |1 > > --- Comment #1 from Richard Biener <rguenth at gcc dot gnu.org> --- > You miss to mark p1 addressable in the alias decl (that is, copy > TREE_ADDRESSABLE). > Do you mean following change: @@ -2334,6 +2334,14 @@ cgraph_node::create_wrapper (cgraph_node *target) cgraph_edge *e = create_edge (target, NULL, 0, CGRAPH_FREQ_BASE); + tree arguments = DECL_ARGUMENTS (decl); + + while (arguments) + { + TREE_ADDRESSABLE (arguments) = false; + arguments = TREE_CHAIN (arguments); + } + expand_thunk (false, true); e->call_stmt_cannot_inline_p = true; Thanks, Martin
diff --git a/gcc/cgraphunit.c b/gcc/cgraphunit.c index 6f61f5c..89c96dc 100644 --- a/gcc/cgraphunit.c +++ b/gcc/cgraphunit.c @@ -2342,6 +2342,14 @@ cgraph_node::create_wrapper (cgraph_node *target) cgraph_edge *e = create_edge (target, NULL, 0, CGRAPH_FREQ_BASE); + tree arguments = DECL_ARGUMENTS (decl); + + while (arguments) + { + TREE_ADDRESSABLE (arguments) = false; + arguments = TREE_CHAIN (arguments); + } + expand_thunk (false, true); e->call_stmt_cannot_inline_p = true; Following patch fixed the issue, boostrap works without any regression. But I guess it's just a partial fix. There's very similar issue: PR63573. Thanks, Martin
Hmm, I think we may get confused by having ADDRESSABLE flag set because the original function body took its address and it makes sense to clear it as Martin suggest. I am bit confused on why this leads to ICE - I think false positive on ADDRESSABLE flag ought to make cgraphunit to introduce extra copy in: if (nargs) for (i = 1, arg = DECL_CHAIN (a); i < nargs; i++, arg = DECL_CHAIN (arg)) { tree tmp = arg; if (!is_gimple_val (arg)) { tmp = create_tmp_reg (TYPE_MAIN_VARIANT (TREE_TYPE (arg)), "arg"); gimple stmt = gimple_build_assign (tmp, arg); gsi_insert_after (&bsi, stmt, GSI_NEW_STMT); } vargs.quick_push (tmp); } that may later end up on stack and prevent call from being a tail call. I see that later optimizers may not be able to optimize out the extra copy simply because it is bit hard to do: one needs to verify that the original location is dead and possibly may be rewritten by a callee. I did not find any code that would prevent clearning ADDRESSABLE flag on parameters, so I think Martin's patch is correct. Additionally I think it is not safe to set TAILCALL flag if the code above introduced a local variable. This is becuase setting this flag by hand bypasses logic in suitable_for_tail_call_opt_p that prevents tail call conversion when any ADDRESSABLE vars exists. I did not find any reasons why ADDRESSABLE flag should be set on parameter besides the usual case that it has address taken, so I tink Martin's patch is OK and correct fix. Martin, does it fix the other PR too?
*** Bug 63721 has been marked as a duplicate of this bug. ***
Author: marxin Date: Fri Nov 7 13:37:41 2014 New Revision: 217222 URL: https://gcc.gnu.org/viewcvs?rev=217222&root=gcc&view=rev Log: PR ipa/63580 * cgraphunit.c (cgraph_node::create_wrapper): TREE_ADDRESSABLE is set to false for a newly created thunk. * g++.dg/ipa/pr63580.C: New test. Added: trunk/gcc/testsuite/g++.dg/ipa/pr63580.C Modified: trunk/gcc/ChangeLog trunk/gcc/cgraphunit.c trunk/gcc/testsuite/ChangeLog
The original testcase still ICEs on ppc64.
(In reply to Markus Trippelsdorf from comment #8) > The original testcase still ICEs on ppc64. Sorry, it actually works fine now.