[CI] Enable assertions for CI

[LyricZhao]

Enable assertions by a new CMake build type TRITON_REL_BUILD_WITH_ASSERTS.

[ptillet]

I don't think we should switch to debug mode in the CI. This will slow down compilation too much and make it untractable. is RelWithDebInfo somehow insufficient?

[LyricZhao]

@ptillet @Superjomn
Hi, the problem is due to the commit 9f27468, at line https://github.com/openai/triton/blob/9f27468377b3dbf1b702c26f5abcfb557b1818e0/lib/Conversion/TritonToTritonGPU/TritonToTritonGPU.cpp#L226.

The commit passed CI but however it can not compile on my own machine, with an error (also seen as current CI results):

/home/phil/.triton/llvm/llvm+mlir-14.0.0-x86_64-linux-gnu-ubuntu-18.04-assert/include/llvm/Support/Casting.h:58:23: error: ‘classof’ is not a member of ‘mlir::TritonGPUTypeConverter’
       58 |     return To::classof(&Val);
          |            ~~~~~~~~~~~^~~~~~

This is because llvm::cast is implemented as:

template <typename To, typename From>
 [[nodiscard]] inline decltype(auto) cast(From *Val) {
   assert(isa<To>(Val) && "cast<Ty>() argument of incompatible type!");
   return CastInfo<To, From *>::doCast(Val);
 }

, while isa requires TritonGPUTypeConverter to implement classof.

But the problem is that the commit can pass CI because RelWithDebInfo disables all asserts (not only this one). This PR is meant to open all asserts for CI to ensure correctness for every commit.

[LyricZhao]

Another thing worth to mention is that the ASSERT_ENABLED_LLVM only have asserts in compiled library, while cast is in a header file that can be ignore by a RelWithDebInfo compiler.

I suggest that we at least align the assertion behaviors.

[ptillet]

it seems like the solution would be to define #NDEBUG rather than switch to full-blown Debug mode. Would this work for you?

[LyricZhao]

it seems like the solution would be to define #NDEBUG rather than switch to full-blown Debug mode. Would this work for you?

LGTM, I'll change to that.

[Superjomn]

Is this bypass the classof check? Seems not safe.

[ptillet]

Same comment here. Isn't cast safer than static_cast here?

[LyricZhao]

@ptillet
Yes, I agree. I've changed to auto typeConverter = getTypeConverter<TritonGPUTypeConverter>(), which leverages the MLIR interface and looks better.

The official implement of this interface is:

   template <typename ConverterTy>
   std::enable_if_t<std::is_base_of<TypeConverter, ConverterTy>::value,
                    ConverterTy *>
   getTypeConverter() const {
     return static_cast<ConverterTy *>(typeConverter);
   }

, which is also static_cast but safely checks using std::is_base_of.