Kim, thank you for the suggestions. Here is an updated webrev:
http://cr.openjdk.java.net/~avoitylov/webrev.8231612.04/ The assembly generated with inline functions is equivalent to that with the macros, and the testing passed as well. -Aleksei On 07/12/2019 02:54, Kim Barrett wrote: >> On Dec 6, 2019, at 8:32 AM, Aleksei Voitylov <[email protected]> >> wrote: >> >> Hi Kim, >> >> (cced hotspot-runtime as the updated webrev has code changes). >> >> Thank you for the suggestion to address this at the code level. The >> following change makes GCC generate correct code: >> >> webrev: http://cr.openjdk.java.net/~avoitylov/webrev.8231612.03/ >> issue: https://bugs.openjdk.java.net/browse/JDK-8231612 >> >> GCC alias analysis at RTL level has some limitations [1]. In particular, >> it sometimes cannot differentiate between pointer and integer and can >> make incorrect assumptions about whether two pointers actually alias >> each other. We are hitting this in >> Atomic::CmpxchgByteUsingInt::operator(), which is used on some platforms. >> >> The solution is to reference the same memory region using only one >> pointer. Anything with a second pointer here continues to confuse GCC, >> and would anyway be fragile in view of this GCC bug which surfaces >> easily on ARM32. >> >> I considered making get/set byte operations inline functions but >> preferred locality defines bring. >> >> Tested with JTReg, JCK and jcstress on ARM32 and Solaris SPARC with no >> regressions. The original problem also disappeared. >> No regressions in JCK VM,Lang, JTReg HotSpot on Linux x86, x86_64, >> AArch64, PPC, Mac. >> >> Thanks, >> >> -Aleksei >> >> [1] https://gcc.gnu.org/bugzilla/show_bug.cgi?id=49330 >> >> On 19/11/2019 19:01, Kim Barrett wrote: >>>> On Nov 19, 2019, at 10:18 AM, Aleksei Voitylov >>>> <[email protected]> wrote: >>>> >>>> Kim, >>>> >>>> you are right, to play it safe we need to -fno-tree-pre all relevant >>>> bool occurrances. I'll get back with an updated webrev when testing is >>>> finished. >>> I think just sprinkling the build system with -fno-tree-pre to address this >>> doesn’t >>> scale in the face of people writing new code. >>> >>> I think we need to find a way to write CmpxchgByteUsingInt in such a way >>> that >>> it doesn’t tickle this bug, or forbid byte-sized cmpxchg (which would be >>> painful), >>> or provide some entirely other implementation approach. > This looks like a good solution to me. > > (I like this approach even without the gcc bug (which was an > entertaining read itself); this change lets me delete an item from my > followup list from templatizing Atomic.) > > A couple of really minor nits: > > (1) In the macros there are various expressions of the form > > x << BitsPerByte * idx > > Please put parens around the shift calculation, e.g. > > x << (BitsPerByte * idx) > > so readers don't need to remember the operator precedence order. > > (2) In this line: > > 871 uint32_t cur = SET_BYTE_IN_INT(*aligned_dest, canon_compare_value, > idx); > > I'd like "*aligned_dest" replaced with "Atomic::load(aligned_dest)". > > In addition, I think functions would be preferable to macros for the > helpers. Put them in the Atomic::CmpxchgByteUsingInt class, with > inline definitions where you put the macro definitions. > >
