[llvm-bugs] [Bug 40280] New: FP exception on optimized code

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=40280

            Bug ID: 40280
           Summary: FP exception on optimized code
           Product: clang
           Version: 6.0
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: LLVM Codegen
          Assignee: unassignedclangbugs at nondot.org
          Reporter: graeser at mi.fu-berlin.de
                CC: llvm-bugs at lists.llvm.org, neeilans at live.com,
                    richard-llvm at metafoo.co.uk

Created attachment 21310
  --> https://bugs.llvm.org/attachment.cgi?id=21310&action=edit
Minimal example

The following code snippet should never divide by zero. However, if compiled
with '-O3' it leads to a floating point exception when called using a certain
pattern. 
See attachment for a minimal example. 

  for (int i=0; i<=3; i++)can be expected. It seems that clang 

    for (int j=0; j<=3; j++)
      for (int k=0; k<=3; k++)
        if ((m!=i) && (m!=j) && (m!=k))
          z += x*(1.0/(m-k))*(1.0/(m-j))*(1.0/(m-i));

If the trapping of FP errors is not enabled, the correct result is computed. So
it seems that clang generates operations for the ruled out cases. Notice that
the order of the factors in the arithmetic expression as well as the way the
function containing this code is called does matter. 

It seems that clang implicitly makes the assumption that user code does not
track FP errors on '-O3'. Gcc only makes this assumption if we explicitly allow
this using '-O3  -fno-trapping-math' where the same behaviour is observed (but
expected).

The problem shows up on a 64bit ubuntu 18.04 with clang-5.0.1 and clang-6.0.0
but not with clang-4.0.0.
Bug 11356 and bug 26969 look related but are either marked fixed or have been
filed against an old clang-3.7 while this one does not show up before
clang-5.0.

-- 
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/20190110/c94221b6/attachment.html>