LLVM ERROR: Cannot select: 0xffff6c0e10d8: i32 = vecreduce_mul 0xffff6c072990

[parched]

Running x.py test --stage 1 on aarch64-unknown-linux-gnu I get one failed test as shown below. This is the first time I've run the tests on this target some I'm unsure if it's a regression.
Nightly commit af50e38

CC @gnzlbg ?

---- [ui] ui/run-pass/simd/simd-intrinsic-generic-reduction.rs stdout ----                                                
                                                                                                                          
error: test compilation failed although it shouldn't!                                                                     
status: exit code: 1                                                                                                      
command: "/usr/src/myapp/build/aarch64-unknown-linux-gnu/stage1/bin/rustc" "/usr/src/myapp/src/test/ui/run-pass/simd/simd-
intrinsic-generic-reduction.rs" "--target=aarch64-unknown-linux-gnu" "--error-format" "json" "-Zui-testing" "-C" "prefer-d
ynamic" "-o" "/usr/src/myapp/build/aarch64-unknown-linux-gnu/test/ui/run-pass/simd/simd-intrinsic-generic-reduction/a" "-C
rpath" "-O" "-Zunstable-options" "-Lnative=/usr/src/myapp/build/aarch64-unknown-linux-gnu/native/rust-test-helpers" "-L" "
/usr/src/myapp/build/aarch64-unknown-linux-gnu/test/ui/run-pass/simd/simd-intrinsic-generic-reduction/auxiliary" "-A" "unu
sed"                                                                                                                      
stdout:                                                                                                                   
------------------------------------------                                                                                
                                                                                                                          
------------------------------------------                                                                                
stderr:                                                                                                                   
------------------------------------------                                                                                
LLVM ERROR: Cannot select: 0xffff6c0e10d8: i32 = vecreduce_mul 0xffff6c072990                                             
  0xffff6c072990: v4i32,ch = load<(load 16 from constant-pool)> 0xffff6c0c6eb8, 0xffff6c0e1620, undef:i64                 
    0xffff6c0e1620: i64 = AArch64ISD::ADDlow 0xffff6c0727f0, TargetConstantPool:i64<<4 x i32> <i32 1, i32 -2, i32 3, i32 4
>> 0 [TF=34]                                                                                                              
      0xffff6c0727f0: i64 = AArch64ISD::ADRP TargetConstantPool:i64<<4 x i32> <i32 1, i32 -2, i32 3, i32 4>> 0 [TF=1]     
        0xffff6c0728c0: i64 = TargetConstantPool<<4 x i32> <i32 1, i32 -2, i32 3, i32 4>> 0 [TF=1]                        
      0xffff6c0726b8: i64 = TargetConstantPool<<4 x i32> <i32 1, i32 -2, i32 3, i32 4>> 0 [TF=34]                         
    0xffff6c0e1008: i64 = undef                                                                                           
In function: _ZN32simd_intrinsic_generic_reduction4main17h19818a8328796d6fE                                               
                                                                                                                          
------------------------------------------                                                                                
                                                                                                                          
thread '[ui] ui/run-pass/simd/simd-intrinsic-generic-reduction.rs' panicked at 'explicit panic', tools/compiletest/src/run
test.rs:3238:9

[gnzlbg]

I think that's some flavor of https://bugs.llvm.org/show_bug.cgi?id=36796.

On stdsimd/packed_simd we work around this by falling back to scalar code instead, so this shouldn't be used anywhere in the ecosystem on aarch64 anyways.

The "fix" for this is to just ignore this test on aarch64.

[sanxiyn]

Update: LLVM PR36796 is fixed.

[mati865]

It's not merged to Rust's LLVM fork.

[Centril]

@sanxiyn @mati865 How about now?

[hanna-kruppe]

FWIW the commit was released in LLVM 9 which rustc currently uses, so this ought to be fixed.

[Centril]

Seems like we need to add a test in that case.

[hanna-kruppe]

No need for a new test, reverting #57827 (not ignoring that test on AArch64 any longer) should suffice. Although I'm not sure if we have CI for that target.

[nikic]

I'd suggest to wait for the LLVM 10 upgrade, which will address some additional issues. While LLVM 9 fixed the majority of this, there were still issues with ordered and non-fast reductions in some cases.

[JohnTitor]

Triage: #57827 is already "reverted" in 6fae7db#diff-1910500d7497a3476e28b7c81e4954faR5 and we're on LLVM 10 now. I think there's nothing more we could do here? Closing.