Hi,
the ipa-inline-transform ICE is caused by fact that devirt_benefit in
ipa-inline-analysis is able to determine devirtualiation oppurtunity of call to
b3, while the ipa-prop responsible for updating function body after inlining is
not. This is because the later is missing code turning known constant into
binfo.
Bootstrapped/regtested x86_64-linux, will commit this shortly.
Honza
PR tree-optimization/55683
* g++.dg/ipa/devirt-9.C: New testcase.
* ipa-prop.c (try_make_edge_direct_virtual_call): Look into constants
for binfo.
Index: testsuite/g++.dg/ipa/devirt-9.C
===================================================================
*** testsuite/g++.dg/ipa/devirt-9.C (revision 0)
--- testsuite/g++.dg/ipa/devirt-9.C (revision 0)
***************
*** 0 ****
--- 1,30 ----
+ /* { dg-do compile } */
+ /* { dg-options "-O2 -fdump-ia-inline" } */
+ double foo ();
+ struct B
+ {
+ bool b1 () { return b3 (); }
+ void b2 ();
+ virtual bool b3 ();
+ };
+ struct C
+ {
+ C () {}
+ bool
+ c1 (float x, float y)
+ {
+ if (x != c3 || y != c4)
+ c2.b2 ();
+ return c2.b1 ();
+ }
+ B c2;
+ float c3, c4;
+ };
+
+ void
+ bar ()
+ {
+ static C c;
+ c.c1 (60, (int) foo ());
+ }
+ /* { dg-final { scan-ipa-dump "Discovered a virtual call to a known target"
"inline" } } */
Index: ipa-prop.c
===================================================================
*** ipa-prop.c (revision 194584)
--- ipa-prop.c (working copy)
*************** try_make_edge_direct_virtual_call (struc
*** 2223,2231 ****
binfo = ipa_value_from_jfunc (new_root_info, jfunc);
! if (!binfo || TREE_CODE (binfo) != TREE_BINFO)
return NULL;
binfo = get_binfo_at_offset (binfo, ie->indirect_info->offset,
ie->indirect_info->otr_type);
if (binfo)
--- 2223,2238 ----
binfo = ipa_value_from_jfunc (new_root_info, jfunc);
! if (!binfo)
return NULL;
+ if (TREE_CODE (binfo) != TREE_BINFO)
+ {
+ binfo = gimple_extract_devirt_binfo_from_cst (binfo);
+ if (!binfo)
+ return NULL;
+ }
+
binfo = get_binfo_at_offset (binfo, ie->indirect_info->offset,
ie->indirect_info->otr_type);
if (binfo)