[llvm-bugs] [Bug 31096] New: Multi-level pointers not disambiguated, even with strict aliasing

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=31096

            Bug ID: 31096
           Summary: Multi-level pointers not disambiguated, even with
                    strict aliasing
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: All
            Status: NEW
          Severity: normal
          Priority: P
         Component: Global Analyses
          Assignee: unassignedbugs at nondot.org
          Reporter: dberlin at dberlin.org
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

The following simple example:
int *p;
int foo(int argc, char **argv)
{
  int result;
  *p = 2;
  if (argc)
    *p = 2;
  result = *p;
  return result;
}

Does not get optimized by GVN (or NewGVN).

It's trivial friend:

int foo(int argc, char **argv, int *p)
{
  int result;
  *p = 2;
  if (argc)
    *p = 2;
  result = *p;
  return result;
}

Does get optimized by NewGVN with a patch i have.

The difference is that the global object turns into an i32 ** that we do a load
from, so the llvm IR looks like this:
 @p = common global i32* null, align 8

 ; Function Attrs: norecurse nounwind ssp uwtable
 define i32 @foo(i32, i8** nocapture readnone) #0 {
   %3 = load i32*, i32** @p, align 8, !tbaa !2
   store i32 2, i32* %3, align 4, !tbaa !6
   %4 = icmp eq i32 %0, 0
   br i1 %4, label %7, label %5

 ; <label>:5                                       ; preds = %2
   %6 = load i32*, i32** @p, align 8, !tbaa !2
   store i32 2, i32* %6, align 4, !tbaa !6
   br label %7

 ; <label>:7                                       ; preds = %2, %5
   %8 = load i32*, i32** @p, align 8, !tbaa !2
   %9 = load i32, i32* %8, align 4, !tbaa !6
   ret i32 %9
 }


 !0 = !{i32 1, !"PIC Level", i32 2}
 !1 = !{!"Apple LLVM version 8.0.0 (clang-800.0.42.1)"}
 !2 = !{!3, !3, i64 0}
 !3 = !{!"any pointer", !4, i64 0}
 !4 = !{!"omnipotent char", !5, i64 0}
 !5 = !{!"Simple C/C++ TBAA"}
 !6 = !{!7, !7, i64 0}
 !7 = !{!"int", !4, i64 0}


Note that in this two level version, we believe the stores to i32* can affect
an i32**.

At the "llvm level" this is correct (IE there are no real types, aso we can't
say anything) , however, with strict aliasing, and the original code, it's not
possible.

In fact, at least for basic types, with strict aliasing, it's not possible for
a pointer of level n to affect a pointer of level n+1 (IE a store to an int *
can't ever alias an int **, only the reverse) , but we haven't taught LLVM
that.

(I suspect, btw, the above can be reproduced with multi-level pointers in
general)

Not sure where we should fix this, suggestions welcome.

-- 
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/20161121/5ec445a5/attachment.html>