This is the mail archive of the gcc-bugs@gcc.gnu.org mailing list for the GCC project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

[Bug tree-optimization/82604] [8 Regression] SPEC CPU2006 410.bwaves ~50% performance regression with trunk@253679 when ftree-parallelize-loops is used


https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82604

--- Comment #4 from Richard Biener <rguenth at gcc dot gnu.org> ---
So it still parallelizes the loop(s) but at one level deeper (line 176 vs.
173).
This is because dependence analysis does not handle calls and loop distribution
distributed a memset.  ISL dependence analysis will have the same issue.

I've quickly tried replacing memset with *p = {} which somewhat works
when wrapping the LHS with a WITH_SIZE_EXPR.  It later ICEs, but well.
RTL expansion suggests we expect it on the RHS, so *p = WITH_SIZE_EXPR <{}, n}
but that doesn't parallelize again (dependence analysis is confused).

I think the proper fix is to dependence analysis in get_references_in_stmt,
similar to how we handle masked loads/stores.

Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]