[llvm-bugs] [Bug 25585] New: Instcombine can't remove a PHI cycle when that cycle should be replaced with a PHI

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=25585

            Bug ID: 25585
           Summary: Instcombine can't remove a PHI cycle when that cycle
                    should be replaced with a PHI
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Scalar Optimizations
          Assignee: unassignedbugs at nondot.org
          Reporter: silviu.baranga at arm.com
                CC: llvm-bugs at lists.llvm.org
    Classification: Unclassified

Created attachment 15319
  --> https://llvm.org/bugs/attachment.cgi?id=15319&action=edit
The reproducer

For the following IR, instcombine should be able to replace %p1 and %p2 with
%p0.

target triple = "aarch64-linux-gnueabi"

define i32 @f(i1 %c) {
entry:
  br i1 %c, label %if.else, label %pre

if.else:
  br label %pre

pre:
  %p0 = phi i32 [ 0, %entry ], [ 1, %if.else ]
  br label %loop

loop:
  %p1 = phi i32 [ %p2, %loop ], [ %p0, %pre ]
  %p2 = phi i32 [ %p0, %loop ] , [ %p1, %pre ]
  br i1 %c, label %loop, label %exit

exit:
  ret i32 %p2
}

We're already trying to do this, but the current phi cycle elimination
algorithm can't handle the case where the value that we should be replacing
other the PHIs with is another PHI (in this case %p).

When the value that we would replace the PHI cycle with is not a PHI, this
algorithm works as expected:

define i32 @g(i1 %c, i32 %val) {
entry:
  br label %pre

pre:
  br label %loop

loop:
  %p1 = phi i32 [ %p2, %loop ], [ %val, %pre ]
  %p2 = phi i32 [ %val, %loop ] , [ %p1, %pre ]
  br i1 %c, label %loop, label %exit

exit:
  ret i32 %p2
}

Reproduce with: opt -instcombine ind.ll -S -o -

-- 
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/20151120/4fd5267b/attachment.html>