[LLVMbugs] [Bug 2003] New: Incorrect iteration count for loop with unsigned ICMP condition

[bugzilla-daemon at cs.uiuc.edu]

http://llvm.org/bugs/show_bug.cgi?id=2003

           Summary: Incorrect iteration count for loop with unsigned ICMP
                    condition
           Product: libraries
           Version: trunk
          Platform: All
        OS/Version: All
            Status: NEW
          Keywords: miscompilation
          Severity: normal
          Priority: P2
         Component: Global Analyses
        AssignedTo: unassignedbugs at nondot.org
        ReportedBy: wmatyjewicz at fastmail.fm
                CC: nicholas at mxc.ca, llvmbugs at cs.uiuc.edu


For the following LLVM code:

define void @foo(i32 %n) {
entry:
        br label %header
header:
        %i = phi i32 [ 100, %entry ], [ %i.inc, %next ]
        %cond = icmp ult i32 %i, %n
        br i1 %cond, label %next, label %return
next:
        %i.inc = add i32 %i, 1
        br label %header
return:
        ret void
}

which contains loop of this form:

unsigned n = ...;
for (unsigned i = 100; i < n; ++i)
  ;

scalar evolution determines loop iteration count as: (-100 + %n).
This isn't correct, because for %n < 100 we'll get negative number of
iterations.

One way to fix it is to add a 'umax' SCEV similar to the 'smax' one. Using it,
the answer for the example would be: (100 umax %n) - 100.

Any other ideas?


-- 
Configure bugmail: http://llvm.org/bugs/userprefs.cgi?tab=email
------- You are receiving this mail because: -------
You are on the CC list for the bug.