[LLVMbugs] [Bug 3831] New: miscompilation with loop index splitting and shorts

[bugzilla-daemon at cs.uiuc.edu]

http://llvm.org/bugs/show_bug.cgi?id=3831

           Summary: miscompilation with loop index splitting and shorts
           Product: libraries
           Version: 2.5
          Platform: PC
        OS/Version: Windows XP
            Status: NEW
          Severity: normal
          Priority: P2
         Component: Loop Optimizer
        AssignedTo: unassignedbugs at nondot.org
        ReportedBy: jay.foad at antixlabs.com
                CC: llvmbugs at cs.uiuc.edu


Created an attachment (id=2724)
 --> (http://llvm.org/bugs/attachment.cgi?id=2724)
source file

With the attached source file I get:

$ llvm-gcc -emit-llvm -c loop.c ; lli loop.o ; echo $?
0

which is correct: f() returns 0 because there is no j in [0..q] such that p + j
== i (with i = 1, q = 2, p = 3).

At -O2 I get:

$ llvm-gcc -O2 -emit-llvm -c loop.c ; lli loop.o ; echo $?
3

which is wrong. The disassembly of f() (inlined into g()) is:

define i32 @g(i16 zeroext %i) nounwind readonly {
entry:
        %0 = zext i16 %i to i32         ; <i32> [#uses=2]
        %1 = load i32* @q, align 4              ; <i32> [#uses=1]
        %lisplit2 = icmp ugt i32 %0, %1         ; <i1> [#uses=1]
        br i1 %lisplit2, label %bb2.i, label %bb1.i

bb1.i:          ; preds = %entry
        %2 = mul i32 %0, 3              ; <i32> [#uses=1]
        ret i32 %2

bb2.i:          ; preds = %entry
        ret i32 0
}

so it's going to return

  (unsigned int)i > q ? 0 : (unsigned int)i * 3

This seems wrong because now there is absolutely no mention of p!

I assume, because of the %lisplit2 variable, that this is something to do with
loop index splitting. If I replace "short" with "int" throughout the source
file, the bug goes away.


-- 
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.