The attached code demonstrates a test program that will fail when optimized at
-O2 and higher for GCC versions 4.4.1 and later, but only when a particular
procedure is inlined (either explicitly at the request of the user, or
implicitly due to auto-inlining).
% gcc --version | head -1
gcc (GCC) 4.4.1 20090725 (Red Hat 4.4.1-2)
The test case will fail if (__may_alias__) is not added to a particular type:
% gcc -O2 test_alias.c
% ./a.out
Abort
And will pass if the "may alias" attribute is added:
% gcc -DMAY_ALIAS -O2 test_alias.c
% ./a.out
The compiler will issue a warning if -Wall is asserted:
% gcc -Wall -Werror -O2 test_alias.c
cc1: warnings being treated as errors
test_alias.c: In function 'main':
test_alias.c:31: error: dereferencing pointer 'dest' does break strict-aliasing
rules
test_alias.c:29: note: initialized from here
test_alias.c:31: error: dereferencing pointer 'src' does break strict-aliasing
rules
test_alias.c:30: note: initialized from here
Note that this check is only implemented if optimization (-O2 or -O3)
enabled, because the strict aliasing optimization only kicks in at those
levels.
This will issue no diagnostic (and perform no strict aliasing optimizations):
% gcc -Wall -Werror -O1 test_alias.c
When inlining is *explicitly* disabled, no errors are detected, and the
resulting program will pass.
% sed -e 's/inline/__attribute__((__noinline__))/' \
test_alias.c > test_alias_without_inline.c
% gcc -Wall -Werror -O2 test_alias_without_inline.c
% ./a.out
If we simply remove the "inline" modifiers, we get the error back, because the
compiler implicitly inlined the static functions.
% sed -e 's/inline//' test_alias.c > test_alias_without_inline.c
% gcc -Wall -Werror -O2 test_alias_without_inline.c
cc1: warnings being treated as errors
test_alias_without_inline.c: In function 'main':
test_alias_without_inline.c:31: error: dereferencing pointer 'dest' does break
strict-aliasing rules
test_alias_without_inline.c:29: note: initialized from here
test_alias_without_inline.c:31: error: dereferencing pointer 'src' does break
strict-aliasing rules
test_alias_without_inline.c:30: note: initialized from here
I don't know what the expected behavior is here, either for GCC or as dictated
by the C standard. However, note that: when this program is compiled without
optimization it works, but it fails when compiled with optimization, and *no*
warnings are issued. The test will pass if inlining is explicitly disabled,
but will fail when either explicit or implicit inlining is present.
One possible solution might be to always issue (by default) the strict-aliasing
warning if the aliasing arises as the result of inlining (if that is
technically possible). Or, always issue the strict-aliasing warning by default?
--
Summary: Inlining changes semantics when strict aliasing
optimization is enabled
Product: gcc
Version: 4.4.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: c
AssignedTo: unassigned at gcc dot gnu dot org
ReportedBy: gary at intrepid dot com
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=42830
