diff mbox

Prevent GCC from folding inline test functions

Message ID 1440162341-26343-1-git-send-email-lgustavo@codesourcery.com
State New
Headers show

Commit Message

Luis Machado Aug. 21, 2015, 1:05 p.m. UTC
While doing some powerpc Linux tests on a ppc 476 board using GCC 5.2, i
noticed inline-bt.exp, inline-cmds.exp and inline-locals.exp failing.

FAIL: gdb.opt/inline-bt.exp: continue to bar (1)
FAIL: gdb.opt/inline-bt.exp: backtrace from bar (1)
FAIL: gdb.opt/inline-bt.exp: continue to bar (2)
FAIL: gdb.opt/inline-bt.exp: backtrace from bar (2)
FAIL: gdb.opt/inline-bt.exp: continue to bar (3)
FAIL: gdb.opt/inline-bt.exp: backtrace from bar (3)
FAIL: gdb.opt/inline-cmds.exp: continue to bar (1)
FAIL: gdb.opt/inline-cmds.exp: backtrace from bar (1)
FAIL: gdb.opt/inline-cmds.exp: continue to bar (2)
FAIL: gdb.opt/inline-cmds.exp: backtrace from bar (2)
FAIL: gdb.opt/inline-cmds.exp: continue to marker
FAIL: gdb.opt/inline-cmds.exp: backtrace from marker
FAIL: gdb.opt/inline-cmds.exp: step into finish marker
FAIL: gdb.opt/inline-locals.exp: continue to bar (1)
FAIL: gdb.opt/inline-locals.exp: continue to bar (2)
FAIL: gdb.opt/inline-locals.exp: backtrace from bar (2)
FAIL: gdb.opt/inline-locals.exp: continue to bar (3)
FAIL: gdb.opt/inline-locals.exp: backtrace from bar (3)

They failed because the breakpoint supposedly inserted at bar was actually
inserted at noinline.

(gdb) break inline-markers.c:20^M
Breakpoint 2 at 0x1000079c: file gdb/testsuite/gdb.opt/inline-markers.c, line 20.^M
(gdb) continue^M
Continuing.^M
^M
Breakpoint 2, noinline () at gdb/testsuite/gdb.opt/inline-markers.c:35^M
35        inlined_fn (); /* inlined */^M

As we can see, line 20 is really inside bar, not noinline:

18 void bar(void)
19 {
20   x += y; /* set breakpoint 1 here */
21 }

Further investigation shows that this is really due to GCC 5's new
ICF pass (-fipa-icf), now enabled by default at -O2, which folds bar
and marker into noinline, where the call to inlined_fn was inlined.

This breaks the testcase since it expects to stop at specific spots.

I thought about two possible fixes for this issue.

- Disable the ICF pass manually when building the binary (-fno-ipa-icf).

This has the advantage of not having to touch the testcase sources themselves,
but the disadvantage of having to add conditional blocks to test the GCC
version. If we ever change GCC's default, we will have to adjust the
conditional block again to match GCC's behavior.

- Modify the testcase sources to make the identical functions unique.

This solution doesn't touch the testcase itself, but changes the source
code slightly in order to make bar, marker and inlined_fn unique. This
causes GCC's ICF pass to ignore these functions and not fold them into
a common identical function.

I'm good with either of them, but i'm more inclined to go with the second
one.

The attached patch implements this by adding the new global variable z, set
to 0, that gets added in different ways to marker and inlined_fn. Since it
is 0, it doesn't affect any possible value checks that we may wish to do
in the future (we currently only check for values changed by bar).

Ok?

ps: I also noticed GDB doesn't do a great job at stating that the breakpoint
was actually inserted at a different source line than previously requested,
so this sounds like a bug that should be fixed, if it is not just wrong
DWARF information (did not investigate it further).

gdb/testsuite/ChangeLog:

	* gdb.opt/inline-bt.c: New global z.
	* gdb.opt/inline-cmds.c: Likewise.
	* gdb.opt/inline-locals.c: Likewise.
	* gdb.opt/inline-markers.c: New extern global z.
	(marker): Use z.
	(inline_fn): Likewise.
---
 gdb/testsuite/gdb.opt/inline-bt.c      | 1 +
 gdb/testsuite/gdb.opt/inline-cmds.c    | 1 +
 gdb/testsuite/gdb.opt/inline-locals.c  | 1 +
 gdb/testsuite/gdb.opt/inline-markers.c | 6 +++---
 4 files changed, 6 insertions(+), 3 deletions(-)

Comments

Pedro Alves Aug. 21, 2015, 4:03 p.m. UTC | #1
On 08/21/2015 02:05 PM, Luis Machado wrote:

> The attached patch implements this by adding the new global variable z, set
> to 0, that gets added in different ways to marker and inlined_fn. Since it
> is 0, it doesn't affect any possible value checks that we may wish to do
> in the future (we currently only check for values changed by bar).
> 
> Ok?
> 

OK, though you should probably make z volatile as well.
Otherwise, soon enough, gcc with LTO sees that z is always 0:

> +int z = 0;

and then these compile down to the same again, and get folded:

> +  x += y - z;
> +  x += y + z;

Thanks,
Pedro Alves
diff mbox

Patch

diff --git a/gdb/testsuite/gdb.opt/inline-bt.c b/gdb/testsuite/gdb.opt/inline-bt.c
index dc2bd45..baa6fd4 100644
--- a/gdb/testsuite/gdb.opt/inline-bt.c
+++ b/gdb/testsuite/gdb.opt/inline-bt.c
@@ -23,6 +23,7 @@ 
 #endif
 
 int x, y;
+int z = 0;
 volatile int result;
 
 void bar(void);
diff --git a/gdb/testsuite/gdb.opt/inline-cmds.c b/gdb/testsuite/gdb.opt/inline-cmds.c
index 9955720..2569289 100644
--- a/gdb/testsuite/gdb.opt/inline-cmds.c
+++ b/gdb/testsuite/gdb.opt/inline-cmds.c
@@ -23,6 +23,7 @@ 
 #endif
 
 int x, y;
+int z = 0;
 volatile int result;
 
 void bar(void);
diff --git a/gdb/testsuite/gdb.opt/inline-locals.c b/gdb/testsuite/gdb.opt/inline-locals.c
index fc018bf..143f7ff 100644
--- a/gdb/testsuite/gdb.opt/inline-locals.c
+++ b/gdb/testsuite/gdb.opt/inline-locals.c
@@ -23,6 +23,7 @@ 
 #endif
 
 int x, y;
+int z = 0;
 volatile int result;
 volatile int *array_p;
 
diff --git a/gdb/testsuite/gdb.opt/inline-markers.c b/gdb/testsuite/gdb.opt/inline-markers.c
index 46c68ae..5d55033 100644
--- a/gdb/testsuite/gdb.opt/inline-markers.c
+++ b/gdb/testsuite/gdb.opt/inline-markers.c
@@ -13,7 +13,7 @@ 
    You should have received a copy of the GNU General Public License
    along with this program.  If not, see <http://www.gnu.org/licenses/>.  */
 
-extern int x, y;
+extern int x, y, z;
 
 void bar(void)
 {
@@ -22,12 +22,12 @@  void bar(void)
 
 void marker(void)
 {
-  x += y; /* set breakpoint 2 here */
+  x += y - z; /* set breakpoint 2 here */
 }
 
 inline void inlined_fn(void)
 {
-  x += y;
+  x += y + z;
 }
 
 void noinline(void)