Hi!
The ubsan pass makes early uninit warning harder (and often doesn't warn),
because e.g. for enum/bool loads we add instrumentation and the reads from
the fields isn't easily visible to the uninit pass, we have
_7 = &a.a;
_8 = MEM[(_Bool *)_7];
instead etc. As the early uninit warning pass shouldn't change the IL
(except for nowarning flags), I think it should be ok to swap the passes,
first warn and only then do ubsan instrumentation.
Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk (or
just for GCC10)?
2019-02-13 Jakub Jelinek <ja...@redhat.com>
PR middle-end/89284
* passes.def: Swap pass_ubsan and pass_early_warn_uninitialized.
* gcc.dg/ubsan/pr89284.c: New test.
--- gcc/passes.def.jj 2019-01-01 12:37:18.041960447 +0100
+++ gcc/passes.def 2019-02-13 14:45:20.205700845 +0100
@@ -56,8 +56,8 @@ along with GCC; see the file COPYING3.
NEXT_PASS (pass_fixup_cfg);
NEXT_PASS (pass_build_ssa);
NEXT_PASS (pass_warn_nonnull_compare);
- NEXT_PASS (pass_ubsan);
NEXT_PASS (pass_early_warn_uninitialized);
+ NEXT_PASS (pass_ubsan);
NEXT_PASS (pass_nothrow);
NEXT_PASS (pass_rebuild_cgraph_edges);
POP_INSERT_PASSES ()
--- gcc/testsuite/gcc.dg/ubsan/pr89284.c.jj 2019-02-13 15:01:25.666165920
+0100
+++ gcc/testsuite/gcc.dg/ubsan/pr89284.c 2019-02-13 15:01:09.987414305
+0100
@@ -0,0 +1,23 @@
+/* PR middle-end/89284 */
+/* { dg-do compile } */
+/* { dg-options "-fsanitize=undefined -O0 -Wuninitialized" } */
+
+struct A { _Bool a; int i; };
+
+int
+foo (void)
+{
+ struct A a;
+ if (a.i) /* { dg-warning "'a.i' is used uninitialized in this function"
} */
+ return 1;
+ return 0;
+}
+
+int
+bar (void)
+{
+ struct A a;
+ if (a.a) /* { dg-warning "'a.a' is used uninitialized in this function"
} */
+ return 1;
+ return 0;
+}
Jakub
