[llvm-bugs] [Bug 24738] New: Loop vectorizer introduces bad "undef"

[via llvm-bugs]

https://llvm.org/bugs/show_bug.cgi?id=24738

            Bug ID: 24738
           Summary: Loop vectorizer introduces bad "undef"
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: normal
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: mattias.v.eriksson at ericsson.com
                CC: llvm-bugs at lists.llvm.org, nrotem at apple.com
    Classification: Unclassified

Created attachment 14851
  --> https://llvm.org/bugs/attachment.cgi?id=14851&action=edit
Reduced reproducer

I think this is a bug in the fix to Bug 14725.

With the attached file if we run:
opt -S -licm -loop-vectorize  -O1 -o /dev/null bugpoint-reduced-simplified.ll
-print-after-all -print-before-all

The loop vectorizer does this transformation:

*** IR Dump Before Loop Vectorization ***
define void @TEST__MAIN() {
  %_tmp52 = call i16 @foo(i16 0)
  %_tmp53 = load i16*, i16** undef
  br label %bb1

bb1:                                              ; preds = %bb1, %0
  %i.22.0 = phi i16 [ undef, %0 ], [ %_tmp55, %bb1 ]
  %_tmp55 = add i16 %i.22.0, 1
  %_tmp59 = icmp slt i16 %_tmp55, 40
  br i1 %_tmp59, label %bb1, label %bb4.preheader

bb4.preheader:                                    ; preds = %bb1
  %_tmp52.lcssa = phi i16 [ %_tmp52, %bb1 ]
  store i16 %_tmp52, i16* %_tmp53
  %_tmp67 = load i16*, i16** undef
  br label %bb4

bb4:                                              ; preds = %bb4.preheader,
%bb4
  br i1 false, label %bb4, label %bb6

bb6:                                              ; preds = %bb4
  store i16 %_tmp52.lcssa, i16* %_tmp67
  ret void
}
*** IR Dump After Loop Vectorization ***
define void @TEST__MAIN() {
  %_tmp52 = call i16 @foo(i16 0)
  %_tmp53 = load i16*, i16** undef
  %min.iters.check = icmp ult i32 1, 4
  br i1 %min.iters.check, label %scalar.ph, label %min.iters.checked

min.iters.checked:                                ; preds = %0
  br i1 true, label %middle.block, label %vector.ph

vector.ph:                                        ; preds = %min.iters.checked
  br label %vector.body

vector.body:                                      ; preds = %vector.body,
%vector.ph
  %index = phi i32 [ 0, %vector.ph ], [ %index.next, %vector.body ]
  %1 = trunc i32 %index to i16
  %broadcast.splatinsert = insertelement <4 x i16> undef, i16 %1, i32 0
  %broadcast.splat = shufflevector <4 x i16> %broadcast.splatinsert, <4 x i16>
undef, <4 x i32> zeroinitializer
  %induction = add <4 x i16> %broadcast.splat, <i16 0, i16 1, i16 2, i16 3>
  %2 = add <4 x i16> %induction, <i16 1, i16 1, i16 1, i16 1>
  %3 = icmp slt <4 x i16> %2, <i16 40, i16 40, i16 40, i16 40>
  %index.next = add i32 %index, 4
  %4 = icmp eq i32 %index.next, 0
  br i1 %4, label %middle.block, label %vector.body, !llvm.loop !0

middle.block:                                     ; preds = %vector.body,
%min.iters.checked
  %resume.val = phi i32 [ 0, %min.iters.checked ], [ 0, %vector.body ]
  %trunc.resume.val = phi i16 [ undef, %min.iters.checked ], [ 0, %vector.body
]
  %cmp.n = icmp eq i32 1, %resume.val
  br i1 %cmp.n, label %bb4.preheader, label %scalar.ph

scalar.ph:                                        ; preds = %middle.block, %0
  %bc.resume.val = phi i32 [ %resume.val, %middle.block ], [ 0, %0 ]
  %bc.trunc.resume.val = phi i16 [ %trunc.resume.val, %middle.block ], [ undef,
%0 ]
  br label %bb1

bb1:                                              ; preds = %bb1, %scalar.ph
  %i.22.0 = phi i16 [ %bc.trunc.resume.val, %scalar.ph ], [ %_tmp55, %bb1 ]
  %_tmp55 = add i16 %i.22.0, 1
  %_tmp59 = icmp slt i16 %_tmp55, 40
  br i1 %_tmp59, label %bb1, label %bb4.preheader, !llvm.loop !3

bb4.preheader:                                    ; preds = %middle.block, %bb1
  %_tmp52.lcssa = phi i16 [ %_tmp52, %bb1 ], [ undef, %middle.block ]

The problem is at the last line. The phi-node has a new incoming value that is
undef, but in this case it should be the same as the first incoming value.

The new "undef" is added in fixLCSSAPHIs() in LoopVectorize.cpp.

-- 
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/20150907/ac5226c7/attachment.html>