[llvm-bugs] [Bug 25986] New: The add instruction is incorrectly folded into icmp when there is an overflow by instcombine

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=25986

            Bug ID: 25986
           Summary: The add instruction is incorrectly folded into icmp
                    when there is an overflow by instcombine
           Product: new-bugs
           Version: 3.7
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: new bugs
          Assignee: unassignedbugs at nondot.org
          Reporter: yuanrui.zhang at intel.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

Given the following reduced test:

define i32 @f(i32 %j) {
entry:
  %inc = add nsw i32 %j, 1
  %cmp = icmp sgt i32 %inc, 0
  %conv = zext i1 %cmp to i32
  ret i32 %conv
}

define i32 @main() {
entry:
  %call = call i32 @f(i32 2147483647)
  ret i32 0
}

InstCombine folds the following 2 instructions

  %inc = add nsw i32 %j, 1
  %cmp = icmp sgt i32 %inc, 0

into:

 <badref> = icmp sgt i32 %j, -1

and the IR becomes wrong after InstCombine as below:


define i32 @f(i32 %j) {
entry:
  %j.lobit = lshr i32 %j, 31
  %j.lobit.not = xor i32 %j.lobit, 1
  ret i32 %j.lobit.not
}

define i32 @main() {
entry:
  %call = call i32 @f(i32 2147483647)
  ret i32 0
}

The error comes from visitICmpInst

-- 
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/20160101/5dc39689/attachment.html>