ICE: void llvm::Value::replaceAllUsesWith(llvm::Value*): Assertion `!contains(New, this) && "this->replaceAllUsesWith(expr(this)) is NOT valid!"' failed.

[bluss]

This is a recently introduced ICE that appears with -Copt-level=3

using rustc 1.13.0-nightly (5531c31 2016-09-12)

rustc: /buildslave/rust-buildbot/slave/nightly-dist-rustc-linux/build/src/llvm/lib/IR/Value.cpp:373: void llvm::Value::replaceAllUsesWith(llvm::Value*): Assertion `!contains(New, this) && "this->replaceAllUsesWith(expr(this)) is NOT valid!"' failed.

Shrunk test case below (original is in ndarray tests/dimension.rs). Use rustc -Copt-level=3 to reproduce.

fn main() {
    remove_axis(&3, 0);
}

trait Dimension {
    fn slice(&self) -> &[usize];
}

impl Dimension for () {
    fn slice(&self) -> &[usize] { &[] }
}

impl Dimension for usize {
    fn slice(&self) -> &[usize] {
        unsafe {
            ::std::slice::from_raw_parts(self, 1)
        }
    }
}

fn remove_axis(value: &usize, axis: usize) -> () {
    let tup = ();
    let mut it = tup.slice().iter();
    for (i, _) in value.slice().iter().enumerate() {
        if i == axis {
            continue;
        }
        it.next();
    }
}

(Edits: Reduced test case further)

[TimNN]

Introduced between nightly-2016-08-03 and nightly-2016-08-04 (also occurs on opt-level 2 here) (Changes)

#35174 has lots of llvm changes, so I guess that's a likely candidate to have caused this.

Edit: Just checked and -Z orbit=off does not reproduce, so I rather suspect #35197

[alexcrichton]

cc @eddyb

[nagisa]

• Assertion happens inside LLVM pass (instcombine?);
• Does not reproduce with LLVM 3.8;

This is the IR which causes the assertion to fire:

; Function Attrs: nounwind readnone uwtable
define internal fastcc void @_ZN4test11remove_axis17h518617d2aa05c61eE() unnamed_addr #1 personality i32 (i32, i32, i64, %"8.unwind::libunwind::_Unwind_Exception"*, %"8.unwind::libunwind::_Unwind_Context"*)* @rust_eh_personality {
entry-block:
  br label %bb7

bb7:                                              ; preds = %bb10.i, %bb12, %bb10, %entry-block
  %iter.sroa.0.0 = phi i64* [ @ref8151, %entry-block ], [ %iter.sroa.0.1, %bb10 ], [ %iter.sroa.0.1, %bb12 ], [ %iter.sroa.0.1, %bb10.i ]
  %iter.sroa.7.0 = phi i64 [ 0, %entry-block ], [ %iter.sroa.7.1, %bb10 ], [ %iter.sroa.7.1, %bb12 ], [ %iter.sroa.7.1, %bb10.i ]
  %it.sroa.0.0 = phi i64* [ getelementptr inbounds ([0 x i64], [0 x i64]* @ref8152, i64 0, i64 0), %entry-block ], [ %it.sroa.0.0, %bb10 ], [ getelementptr inbounds ([0 x i64], [0 x i64]* @ref8152, i64 0, i64 0), %bb12 ], [ %7, %bb10.i ]
  %tmp10.sroa.0.0 = phi i64 [ undef, %entry-block ], [ 0, %bb10 ], [ %tmp10.sroa.0.1, %bb12 ], [ %tmp10.sroa.0.1, %bb10.i ]
  %0 = icmp eq i64* %iter.sroa.0.0, getelementptr inbounds (i64, i64* @ref8151, i64 1)
  br i1 %0, label %"_ZN81_$LT$core..iter..Enumerate$LT$I$GT$$u20$as$u20$core..iter..iterator..Iterator$GT$4next17h32f5bf83064acb8cE.exit", label %"_ZN91_$LT$core..slice..Iter$LT$$u27$a$C$$u20$T$GT$$u20$as$u20$core..iter..iterator..Iterator$GT$4next17hbcf540387e7bb34eE.exit.i"

"_ZN91_$LT$core..slice..Iter$LT$$u27$a$C$$u20$T$GT$$u20$as$u20$core..iter..iterator..Iterator$GT$4next17hbcf540387e7bb34eE.exit.i": ; preds = %bb7
  %1 = getelementptr inbounds i64, i64* %iter.sroa.0.0, i64 1
  %2 = add i64 %iter.sroa.7.0, 1
  %3 = ptrtoint i64* %iter.sroa.0.0 to i64
  br label %"_ZN81_$LT$core..iter..Enumerate$LT$I$GT$$u20$as$u20$core..iter..iterator..Iterator$GT$4next17h32f5bf83064acb8cE.exit"

"_ZN81_$LT$core..iter..Enumerate$LT$I$GT$$u20$as$u20$core..iter..iterator..Iterator$GT$4next17h32f5bf83064acb8cE.exit": ; preds = %bb7, %"_ZN91_$LT$core..slice..Iter$LT$$u27$a$C$$u20$T$GT$$u20$as$u20$core..iter..iterator..Iterator$GT$4next17hbcf540387e7bb34eE.exit.i"
  %iter.sroa.0.1 = phi i64* [ %1, %"_ZN91_$LT$core..slice..Iter$LT$$u27$a$C$$u20$T$GT$$u20$as$u20$core..iter..iterator..Iterator$GT$4next17hbcf540387e7bb34eE.exit.i" ], [ getelementptr inbounds (i64, i64* @ref8151, i64 1), %bb7 ]
  %iter.sroa.7.1 = phi i64 [ %2, %"_ZN91_$LT$core..slice..Iter$LT$$u27$a$C$$u20$T$GT$$u20$as$u20$core..iter..iterator..Iterator$GT$4next17hbcf540387e7bb34eE.exit.i" ], [ %iter.sroa.7.0, %bb7 ]
  %tmp10.sroa.6.0 = phi i64 [ %3, %"_ZN91_$LT$core..slice..Iter$LT$$u27$a$C$$u20$T$GT$$u20$as$u20$core..iter..iterator..Iterator$GT$4next17hbcf540387e7bb34eE.exit.i" ], [ 0, %bb7 ]
  %tmp10.sroa.0.1 = phi i64 [ %iter.sroa.7.0, %"_ZN91_$LT$core..slice..Iter$LT$$u27$a$C$$u20$T$GT$$u20$as$u20$core..iter..iterator..Iterator$GT$4next17hbcf540387e7bb34eE.exit.i" ], [ %tmp10.sroa.0.0, %bb7 ]
  %4 = inttoptr i64 %tmp10.sroa.6.0 to i64*
  %switchtmp = icmp eq i64* %4, null
  br i1 %switchtmp, label %bb9, label %bb10

bb9:                                              ; preds = %"_ZN81_$LT$core..iter..Enumerate$LT$I$GT$$u20$as$u20$core..iter..iterator..Iterator$GT$4next17h32f5bf83064acb8cE.exit"
  ret void

bb10:                                             ; preds = %"_ZN81_$LT$core..iter..Enumerate$LT$I$GT$$u20$as$u20$core..iter..iterator..Iterator$GT$4next17h32f5bf83064acb8cE.exit"
  %5 = icmp eq i64 %tmp10.sroa.0.1, 0
  br i1 %5, label %bb7, label %bb12

bb12:                                             ; preds = %bb10
  %6 = icmp eq i64* %it.sroa.0.0, getelementptr inbounds ([0 x i64], [0 x i64]* @ref8152, i64 0, i64 0)
  br i1 %6, label %bb7, label %bb10.i

bb10.i:                                           ; preds = %bb12
  %7 = getelementptr inbounds i64, i64* %it.sroa.0.0, i64 1
  br label %bb7
}

Most likely A-llvm stuff.

[nagisa]

Filled https://llvm.org/bugs/show_bug.cgi?id=30402

[nagisa]

Fixed in https://reviews.llvm.org/rL281650, need to update/patch our LLVM.

[nrc]

Discussed in compiler meeting, we prefer to wait for the LLVM upgrade to fix this. But if it looks like we hit beta without it, we can backport the patch to our LLVM.

triage: p-medium

assign @nagisa in my head, but not on GitHub (because I can't)

[nrc]

Turns out I was too eager - this in fact affects beta, and so we should backport the LLVM patch now.

[nrc]

assigning to myself as a proxy for @nagisa

[bluss]

Confirming that the original ndarray tests/dimension.rs issue was fixed by rustc 1.13.0-nightly (32571c0 2016-09-17)

[sylvestre]

I asked for this patch to be backported in 3.9.1: https://llvm.org/bugs/show_bug.cgi?id=30261#c14