On Mon, Jun 6, 2022 at 2:12 AM Hongbo Liu via Gcc-patches <gcc-patches@gcc.gnu.org> wrote: > > Hi, > > This is my first time using email patch, please correct me if there is any > error. Thanks! > > > I found a c++ optimization bug first introduced via > commit 520d5ad337eaa15860a5a964daf7ca46cf31c029, which will make program > compiled with -O2 have unexpected behavior.
Except your testcase violates C/C++ aliasing rules. You store to offset_delta via int64_t (which is long on x86_64) but then do a load via long long. If you fix the testcase, it works correctly at -O2 and above. If you don't want strict aliasing to be turned on you can use -fno-strict-aliasing. Your patch is wrong as now none of the modref data will be used which is exposing the bug in the testcase and it is a bug in the source code you are compiling. As for how to add the testcase, https://gcc.gnu.org/wiki/HowToPrepareATestcase does have some information. So does https://gcc.gnu.org/onlinedocs/gccint/Testsuites.html#Testsuites . Thanks, Andrew Pinski > > > How to produce > > > 1. Checkout to latest master > commit: c4d702fb3c1e2f6e1bc8711da81bff59543b1b19 > 2. Build and install the gcc; > 3. Compile the attached file `test.cc` with command: `latest-g++ -std=c++20 > -g0 -O1 test.cc -o test_o1`; > 4. The output of `./test_o1` will be `offset delta: 5`; > 5. Compile the attached file `test.cc` with command: `latest-g++ > -std=c++20 -g0 -O2 test.cc -o test_o2`; > 6. The output of `./test_o2` will be `offset delta: 0`; > > > The program behavior with `-O2` is inconsistent with `-O1` and unexpected. > > > How to fix > > > The bug was introduced via commit 520d5ad337eaa15860a5a964daf7ca46cf31c029. > The attached patch file could fix the problem and make `-O2` and `-O1` have > same behavior. > > > Request for suggestion > > > I could not find and document about how to add this kind of test, could > anyone give me some suggestions?
