[llvm-bugs] [Bug 49900] New: Vectorizer is querying SCEV on partially constructed IR

[via llvm-bugs]

https://bugs.llvm.org/show_bug.cgi?id=49900

            Bug ID: 49900
           Summary: Vectorizer is querying SCEV on partially constructed
                    IR
           Product: libraries
           Version: trunk
          Hardware: PC
                OS: Linux
            Status: NEW
          Severity: enhancement
          Priority: P
         Component: Loop Optimizer
          Assignee: unassignedbugs at nondot.org
          Reporter: listmail at philipreames.com
                CC: llvm-bugs at lists.llvm.org

This bug is a reduced form of
https://bugs.chromium.org/p/oss-fuzz/issues/detail?id=33029 which was reported
as a regression against 4ffcd4fe9ac2ee948948f732baa16663eb63f1c7.  I've further
reduced the test case and confirmed that this is not actually related to
multiple exit vectorization.  The test case below involves no multiple or early
exit loops, and disabling the multiple exit support entirely does not change
the outcome.  

$ ./opt -S test.ll -loop-vectorize # crashes
$ cat test.ll
; ModuleID =
'/home/preames/Downloads/clusterfuzz-testcase-minimized-llvm-opt-fuzzer--x86_64-loop_vectorize-5093295945547776'
source_filename =
"llvm-project/llvm/test/Transforms/LoopVectorize/X86/pr39160.ll"
target datalayout =
"e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128-ni:1"
target triple = "x86_64-unknown-linux-gnu"

define void @barney() {
bb:
  br label %bb2

bb2:                                              ; preds = %bb2, %bb
  %tmp4 = icmp slt i32 undef, 0
  br i1 %tmp4, label %bb2, label %bb5

bb5:                                              ; preds = %bb2
  br label %bb19

bb18:                                             ; preds = %bb33
  ret void

bb19:                                             ; preds = %bb36, %bb5
  %tmp22 = phi i32 [ %tmp65, %bb36 ], [ undef, %bb5 ]
  br label %bb50

bb33:                                             ; preds = %bb62
  br i1 undef, label %bb18, label %bb36

bb36:                                             ; preds = %bb33
  br label %bb19

bb46:                                             ; preds = %bb50
  %C = icmp uge i8 16, 0
  %B4 = and i1 %C, true
  %B5 = xor i1 %C, undef
  br i1 %B5, label %bb48, label %bb59

bb48:                                             ; preds = %bb46
  ret void

bb50:                                             ; preds = %bb50, %bb19
  %tmp52 = phi i32 [ %tmp55, %bb50 ], [ %tmp22, %bb19 ]
  %tmp53 = phi i64 [ %tmp56, %bb50 ], [ 1, %bb19 ]
  %tmp54 = add i32 %tmp52, 12
  %tmp55 = add i32 %tmp52, 13
  %tmp56 = add nuw nsw i64 %tmp53, 1
  %C6 = icmp sle i32 %tmp54, 65536
  br i1 %C6, label %bb50, label %bb46

bb59:                                             ; preds = %bb46
  br label %bb62

bb62:                                             ; preds = %bb68, %bb59
  %tmp63 = phi i32 [ %tmp65, %bb62 ], [ %tmp55, %bb59 ]
  %tmp64 = phi i64 [ %tmp66, %bb62 ], [ %tmp56, %bb59 ]
  %tmp65 = add i32 %tmp63, 13
  %tmp66 = add nuw nsw i64 %tmp64, 1
  %C1 = icmp ult i32 %tmp65, 65536
  br i1 %C1, label %bb62, label %bb33
}

We're crashing way down in SCEV when the predecessor lists of a phi and the
basic block itself don't match.  This can be seen in the basic block below:

bb46:                                             ; preds = %middle.block,
%bb50
  %tmp55.lcssa = phi i32 [ %tmp55, %bb50 ]
  %tmp56.lcssa = phi i64 [ %tmp56, %bb50 ]
  %C = icmp uge i8 16, 0
  %B4 = and i1 %C, true
  %B5 = xor i1 %C, undef
  br i1 %B5, label %bb48, label %bb59

This is illegal usage of SCEV by LoopVectorize.  It appears to be a long
standing issue related to how we set up the phis for newly added edges.

-- 
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/20210408/b9bad78c/attachment.html>