[llvm-bugs] [Bug 27869] New: Optimizer doesn't simplify obvious contradiction

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=27869

            Bug ID: 27869
           Summary: Optimizer doesn't simplify obvious contradiction
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Scalar Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: me at manueljacob.de
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

The following C code contains an "obvious" contradiction.  An integer can't
both be 0 and negative, so this should be simplifies to false:

_Bool test(int i) {
  return (i == 0) & (i < 0);
}

The LLVM IR with SROA run on it:

define zeroext i1 @test(i32 %i) #0 {
entry:
  %cmp = icmp eq i32 %i, 0
  %conv = zext i1 %cmp to i32
  %cmp1 = icmp slt i32 %i, 0
  %conv2 = zext i1 %cmp1 to i32
  %and = and i32 %conv, %conv2
  %tobool = icmp ne i32 %and, 0
  ret i1 %tobool
}

However running all optimizations on it will result in the following code:

define zeroext i1 @test(i32 %i) #0 {
entry:
  %cmp = icmp eq i32 %i, 0
  %conv = zext i1 %cmp to i32
  %i.lobit = lshr i32 %i, 31
  %and = and i32 %conv, %i.lobit
  %tobool = icmp ne i32 %and, 0
  ret i1 %tobool
}

The problem is that InstCombine transforms the (zext (< i 0)) pattern into
clever bit shifting: (lshr i 31).  At first this saves one operation, but then
the obvious contradiction is obfuscated.

I see two options how to solve this:

1) Deferring the (zext (< i 0)) -> (lshr i 31) transformation, allowing
InstCombine to simplify the contradiction.

2) Making subsequent transformations more clever, so they can "look through"
the lshr instruction, seeing that it's actually a "less than" predicate.

-- 
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/20160525/89848fa0/attachment-0001.html>