[llvm-bugs] [Bug 38529] New: Alias analysis fails on aliased function arguments -- possibly resulting in incorrect code after scheduling

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=38529

            Bug ID: 38529
           Summary: Alias analysis fails on aliased function arguments --
                    possibly resulting in incorrect code after scheduling
           Product: new-bugs
           Version: 6.0
          Hardware: Other
                OS: All
            Status: NEW
          Severity: release blocker
          Priority: P
         Component: new bugs
          Assignee: unassignedbugs at nondot.org
          Reporter: ruttenberg at reservoir.com
                CC: llvm-bugs at lists.llvm.org

Created attachment 20679
  --> https://bugs.llvm.org/attachment.cgi?id=20679&action=edit
C program example

Attached in an example program (taken from the gcc test suite).  Note that the
order fo the stores to pi and pl must happen in order because the argument to f
may alias, and in fact in this example they do.  So the machine scheduler will
need a DAG edge from the store to pi to the store to pl.

To see the problem, we need a target that schedules with alias analysis
enabled.  Arm64 cortex-53 is one example of this.  So compile with:

    clang --target=aarch64-arm-none-eabi -mcpu=cortex-a53  -O2 -S store-store.c
-mllvm -debug-only=isel -mllvm -view-misched-dags

I have attached the dot file for the function f that this generates. Note the
lack of an edge from between the STRX and the STRW.  I have also attached the
result store-store.s.  The code generated is clearly wrong.  It performs the
stores in the wrong order and returns 1 instead of 0.

-- 
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/20180811/f530a6b3/attachment.html>