[llvm-bugs] [Bug 47679] New: Miscompile for AArch64, possibly related to llvm.umul.with.overflow.i64

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=47679

            Bug ID: 47679
           Summary: Miscompile for AArch64, possibly related to
                    llvm.umul.with.overflow.i64
           Product: new-bugs
           Version: 11.0
          Hardware: Other
                OS: All
            Status: NEW
          Severity: release blocker
          Priority: P
         Component: new bugs
          Assignee: unassignedbugs at nondot.org
          Reporter: kinke at gmx.net
                CC: htmldeveloper at gmail.com, llvm-bugs at lists.llvm.org

Created attachment 24004
  --> https://bugs.llvm.org/attachment.cgi?id=24004&action=edit
unoptimized input IR

Hi,

I'm seeing test regressions with LLVM 11 (rc3 and rc4) on AArch64 for an
*unoptimized* function performing a i64 multiply-with-overflow-check
(apparently wrongly reporting overflows in some cases). Unfortunately, I don't
have an AArch64 box to test, and no according qemu setup, so this is a bit of a
shot in the dark. I've attached the source IR.

The generated *optimized* machine assembly is identical for LLVM 10 and 11, and
seems to work fine. The *unoptimized* assembly versions differ though; I don't
have a clue about AArch64, but in the optimized case and in the unoptimized 10
version, there's a `cmp` right after the `umulh`, and that's missing in the
unoptimized 11 version...

-- 
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/20200929/0e81b261/attachment.html>