[llvm-bugs] [Bug 39994] New: Converting invoke to call does not transfer "MD_mem_parallel_loop_access" metadata

via llvm-bugs llvm-bugs at lists.llvm.org
Thu Dec 13 05:45:26 PST 2018


https://bugs.llvm.org/show_bug.cgi?id=39994

            Bug ID: 39994
           Summary: Converting invoke to call does not transfer
                    "MD_mem_parallel_loop_access" metadata
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Transformation Utilities
          Assignee: unassignedbugs at nondot.org
          Reporter: moritz.kreutzer at siemens.com
                CC: llvm-bugs at lists.llvm.org

Hi,

We have observed a problem with vectorization and the newest LLVM/Clang trunk
(older versions seem to be affected as well). We have a loop with a potential
race condition for which we force vectorization using "#pragma clang loop
vectorize(assume_safety)." Despite that, Clang fails to vectorize our loop
reporting that it "cannot identify array bounds." After investigating the IR
after each pass, we found a potential cause of this issue. It seems like the
"MD_mem_parallel_loop_access" metadata did not get transferred over when
converting an invoke to a call. We could fix the problem by adding the
following line to the changeToCall() method in Local.cpp:

NewCall->copyMetadata(*II, LLVMContext::MD_mem_parallel_loop_access);

This is similar to to what is done in SROA, but we are not sure whether this is
a valid fix for the observed issue.

Please note that we are aware of the WIP under D52116 and D52117, but our fix
(with additional copying of "LLVMContext::MD_access_group" in the same call)
was also required to get the loop vectorized after applying those two patches.


Thanks,
Moritz

-- 
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20181213/950dddae/attachment.html>


More information about the llvm-bugs mailing list