| Message ID | 20131125110244.GO30062@redhat.com |
|---|---|
| State | New |
| Headers | show |
On Mon, Nov 25, 2013 at 12:02 PM, Marek Polacek <polacek@redhat.com> wrote: > On Mon, Nov 25, 2013 at 11:47:29AM +0100, Jakub Jelinek wrote: >> That looks much better, I wonder if it would reproduce even if the result >> is saved somewhere (either E *e; field in S with swapping of the two class >> definitions, or global var or static data member static E *e; of S), >> otherwise the new just allocates and throws away. Ok with that change, or >> if it makes the problem no longer reproduceable, ok as is. > > It is reproduceable with all three variants. I'll commit this one shortly. > >> Note, once the LTO issues with internal functions are fixed, we need >> to grep for all those dg-skip-if and remove them. > > Yeah, definitely. I'm keeping it in mind. I'll pre-approve removing the assert for ->leaf that currently makes internal fns ICE with LTO for now. If it makes them "work" (please double-check also code generation results). I'll coordinate with Honza on who is going to fix this for real. Richard. > 2013-11-25 Marek Polacek <polacek@redhat.com> > > testsuite/ > * g++.dg/ubsan/pr59250.C: New test. > > --- gcc/testsuite/g++.dg/ubsan/pr59250.C.mp3 2013-11-25 10:43:24.797315678 +0100 > +++ gcc/testsuite/g++.dg/ubsan/pr59250.C 2013-11-25 11:55:37.000000000 +0100 > @@ -0,0 +1,19 @@ > +// PR sanitizer/59250 > +// { dg-do compile } > +// { dg-options "-fsanitize=undefined" } > +// { dg-skip-if "" { *-*-* } { "-flto" } { "" } } > + > +struct E { > + int i; > +}; > + > +struct S { > + const char *s; > + S (const char *); > + static E *e; > +}; > + > +S::S (const char *) : s (0) > +{ > + e = new E (); > +} > > Marek
On Tue, Nov 26, 2013 at 10:13:35AM +0100, Richard Biener wrote: > On Mon, Nov 25, 2013 at 12:02 PM, Marek Polacek <polacek@redhat.com> wrote: > > On Mon, Nov 25, 2013 at 11:47:29AM +0100, Jakub Jelinek wrote: > >> That looks much better, I wonder if it would reproduce even if the result > >> is saved somewhere (either E *e; field in S with swapping of the two class > >> definitions, or global var or static data member static E *e; of S), > >> otherwise the new just allocates and throws away. Ok with that change, or > >> if it makes the problem no longer reproduceable, ok as is. > > > > It is reproduceable with all three variants. I'll commit this one shortly. > > > >> Note, once the LTO issues with internal functions are fixed, we need > >> to grep for all those dg-skip-if and remove them. > > > > Yeah, definitely. I'm keeping it in mind. > > I'll pre-approve removing the assert for ->leaf that currently makes > internal fns ICE with LTO for now. If it makes them "work" (please > double-check also code generation results). Unfortunately, only removing the assert for ->leaf does not work, without it I get: ./a.out [Leaving LTRANS /tmp/ccvr1cAT.args] [Leaving LTRANS /tmp/ccfdPwKs.ltrans.out] [Leaving LTRANS /tmp/cctcbfzL.args] [Leaving LTRANS /tmp/ccfdPwKs.ltrans0.o] /tmp/ccfdPwKs.ltrans0.ltrans.o: In function `main': ccfdPwKs.ltrans0.o:(.text+0x1d): undefined reference to `.Lubsan_data0.2659' (Though it may be a bug in ubsan.) Marek
--- gcc/testsuite/g++.dg/ubsan/pr59250.C.mp3 2013-11-25 10:43:24.797315678 +0100 +++ gcc/testsuite/g++.dg/ubsan/pr59250.C 2013-11-25 11:55:37.000000000 +0100 @@ -0,0 +1,19 @@ +// PR sanitizer/59250 +// { dg-do compile } +// { dg-options "-fsanitize=undefined" } +// { dg-skip-if "" { *-*-* } { "-flto" } { "" } } + +struct E { + int i; +}; + +struct S { + const char *s; + S (const char *); + static E *e; +}; + +S::S (const char *) : s (0) +{ + e = new E (); +}