[PATCH] D91781: [VE] Add regression test for D91151

[David Blaikie via Phabricator via llvm-commits]

dblaikie added a comment.

In D91781#2406553 <https://reviews.llvm.org/D91781#2406553>, @kaz7 wrote:

> In D91781#2406521 <https://reviews.llvm.org/D91781#2406521>, @dblaikie wrote:
>
>> with assertions enabled/if the old code asserting, you may be able to simplify the test case down much further, as the assertion will ensure a more reliable/fast failure rather than some unspecified failure due to undefined behavior in the execution later on/elsewhere.
>
> That's make sens.  However, cast mechanism is compilcated and I couldn't find any good way to simplify this test case.  Anyway, this is an assertion message  and some back trace from old code.
>
>   llc: /home/jam/llvm-upstream/llvm-project/llvm/include/llvm/Support/Casting.h:262: typename llvm::cast_retty<To, From>::ret_type llvm::cast(Y&) [with X = llvm::ConstantSDNode; Y = llvm::SDValue; typename llvm::cast_retty<To, From>::ret_type = llvm::ConstantSDNode*]: Assertion `isa<X>(Val) && "cast<Ty>() argument of incompatible type!"' failed.
>   ...
>    #6 0x00007fbf5a26b859 abort (/lib/x86_64-linux-gnu/libc.so.6+0x25859)
>    #7 0x00007fbf5a26b729 (/lib/x86_64-linux-gnu/libc.so.6+0x25729)
>    #8 0x00007fbf5a27cf36 (/lib/x86_64-linux-gnu/libc.so.6+0x36f36)
>    #9 0x0000559f57608668 llvm::cast_retty<llvm::ConstantSDNode, llvm::SDValue>::ret_type llvm::cast<llvm::ConstantSDNode, llvm::SDValue>(llvm::SDValue&) /home/jam/llvm-upstream/llvm-project/llvm/include/llvm/Support/Casting.h:262:3
>   #10 0x0000559f5792a5af (anonymous namespace)::VEDAGToDAGISel::selectADDRzii(llvm::SDValue, llvm::SDValue&, llvm::SDValue&, llvm::SDValue&) /home/jam/llvm-upstream/llvm-project/llvm/lib/Target/VE/VEISelDAGToDAG.cpp:230:38
>   #11 0x0000559f5792965d (anonymous namespace)::VEDAGToDAGISel::CheckComplexPattern(llvm::SDNode*, llvm::SDNode*, llvm::SDValue, unsigned int, llvm::SmallVectorImpl<std::pair<llvm::SDValue, llvm::SDNode*> >&) /home/jam/llvm-upstream/build-debug/lib/Target/VE/VEGenDAGISel.inc:11266:100

With the use of CReduce <https://embed.cs.utah.edu/creduce/> & some manual reduction, I managed to get the test down to this:

  define i32 @a(i32* %b) {
  if.c:
          %d = load atomic i32, i32* %b monotonic, align 4
          switch i32 %d, label %e.default [
                  i32 1, label %e.bb
                  i32 2, label %e.f
                  i32 4, label %e.g
                  i32 0, label %if.end
          ]
  e.bb:
          br label %if.end
  e.f:
          br label %if.end
  e.g:
          br label %if.end
  e.default:
          br label %if.end
  if.end:
          %ret.0 = phi i32
                  [ 0, %e.default ],
                  [ 6, %if.c ],
                  [ 0, %e.g ],
                  [ 1, %e.f ],
                  [ 1, %e.bb ]
          ret i32 %ret.0
  }

Though it'd be good if you could check this over and ensure this test makes sense for the failure - and see if there's any domain-specific knowledge you could apply to change if it could be modified/simplified/be more explicit about how/why this code is what's needed to reach the buggy case.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D91781/new/

https://reviews.llvm.org/D91781