[llvm-bugs] [Bug 44186] New: Incorrect instcombine transform urem -> icmp+zext with vectors

via llvm-bugs llvm-bugs at lists.llvm.org
Sat Nov 30 07:28:15 PST 2019


https://bugs.llvm.org/show_bug.cgi?id=44186

            Bug ID: 44186
           Summary: Incorrect instcombine transform urem -> icmp+zext with
                    vectors
           Product: libraries
           Version: trunk
          Hardware: All
                OS: All
            Status: NEW
          Keywords: miscompilation
          Severity: normal
          Priority: P
         Component: Scalar Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: nunoplopes at sapo.pt
                CC: lebedev.ri at gmail.com, llvm-bugs at lists.llvm.org,
                    regehr at cs.utah.edu, spatel+llvm at rotateright.com

Testcase from Transforms/InstCombine/vector-urem.ll:

define <4 x i32> @test_v4i32_one_undef(<4 x i32> %a0) {
  %1 = urem <4 x i32> { 1, 1, 1, undef }, %a0
  ret <4 x i32> %1
}
=>
define <4 x i32> @test_v4i32_one_undef(<4 x i32> %a0) {
  %1 = icmp ne <4 x i32> %a0, { 1, 1, 1, undef }
  %2 = zext <4 x i1> %1 to <4 x i32>
  ret <4 x i32> %2
}
Transformation doesn't verify!
ERROR: Value mismatch

Example:
<4 x i32> %a0 = < #x00000010 (16), #x00000001 (1), #x00000400 (1024),
#x00000001 (1) >

Source:
<4 x i32> %1 = < #x00000001 (1), #x00000000 (0), #x00000001 (1), #x00000000 (0)
>

Target:
<4 x i1> %1 = < #x1 (1), #x0 (0), #x1 (1), #x1 (1) >
<4 x i32> %2 = < #x00000001 (1), #x00000000 (0), #x00000001 (1), #x00000001 (1)
>
Source value: < #x00000001 (1), #x00000000 (0), #x00000001 (1), #x00000000 (0)
>
Target value: < #x00000001 (1), #x00000000 (0), #x00000001 (1), #x00000001 (1)
>


Essentially, "x urem 1" is 0 for any 'x'. However, the optimized code can
produce 0 or 1.
I guess the easiest fix is to change undef to 1 in the icmp.

-- 
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/20191130/7105e713/attachment.html>


More information about the llvm-bugs mailing list