[Arith] MLIR PresburgerSet compile fix mlir >= 160

[cbalint13]

Hi folks,

Some fixes for MLIR based analyzer module introduced by #14690 .

---

• Make CMake at par with LLVM info:

{...}
-- Use llvm-config=llvm-config-64
-- LLVM libdir: /usr/lib64
-- Found MLIR
-- Build with MLIR
-- Set TVM_MLIR_VERSION=160
-- Found LLVM_INCLUDE_DIRS=/usr/include
{...}
--    USE_MKL                            : OFF
--    USE_MLIR                           : ON
--    USE_MSVC_MT                        : OFF
{...}

• Fix several compilation errors:

error: cannot convert 'llvm::SmallVector<long int>' to 'llvm::ArrayRef<mlir::presburger::MPInt>'
error: no matching function for call to 'tvm::IntImm::IntImm(tvm::runtime::DataType, mlir::presburger::MPInt&)'
note:   no known conversion for argument 2 from 'mlir::presburger::MPInt' to 'int64_t' {aka 'long int'}

Tested using: llvm/mlir 16.0.6, llvm/mlir 15.0.7, llvm/mlir 17.0.0rc3

Thanks,
~Cristian.

Cc @multiverstack-intellif , @wrongtest-intellif , @vinx13, @Hzfengsy , @junrushao , @tqchen

[junrushao]

Thanks for the bugfixes! Definitely not an expert here, and do you think we need to add a test for changes introduced in presurger_set.cc?

[cbalint13]

Thanks for the bugfixes! Definitely not an expert here, and do you think we need to add a test for changes introduced in presurger_set.cc?

@junrushao ,

Me either, not expert on this topic, unsure how #14690 passed it's own test (can't see related CI logs).
Since here we have compilation fixes (no changes), the #14690 test case included should be sufficient.

---

Let's wait the tests to pass in green, reported issue here might be related to:
(a) a different mlir = 16 version on my side
(b) a newer gcc 13 not willing to infer casts (unplausible)

If (a) will adjust this PR with some little #if version to keep pace with MLIR API changes (mlir >= 15, as expected).

Let's also ask authors: @multiverstack-intellif , @wrongtest-intellif

[cbalint13]

If (a) will adjust this PR with some little #if version to keep pace with MLIR API changes (mlir >= 15, as expected).
Let's also ask authors: @multiverstack-intellif , @wrongtest-intellif

PR is now updated to be consistent with all mlir >= 15 releases.
Tested for: llvm/mlir 15.0.7, llvm/mlir 16.0.6, llvm/mlir 17.0.0rc3.

[multiverstack-intellif]

Thanks for the bugfixes! Definitely not an expert here, and do you think we need to add a test for changes introduced in presurger_set.cc?

@junrushao ,

Me either, not expert on this topic, unsure how #14690 passed it's own test (can't see related CI logs). Since here we have compilation fixes (no changes), the #14690 test case included should be sufficient.

Let's wait the tests to pass in green, reported issue here might be related to: (a) a different mlir = 16 version on my side (b) a newer gcc 13 not willing to infer casts (unplausible)

If (a) will adjust this PR with some little #if version to keep pace with MLIR API changes (mlir >= 15, as expected).

Let's also ask authors: @multiverstack-intellif , @wrongtest-intellif

I believe it's because of a), USE_MLIR is not turned ON in CI test, AFAIK, and only LLVM 14 & 15 is tested locally when #14690 is submitted.

[junrushao]

Sounds great, and it seems that we all agree with the changes being made! Will merge it in as soon as the CI is green.

[cbalint13]

@multiverstack-intellif , @junrushao

Trying to tackle similar issue as described at interblock analisys I noticed it can't compile on newer mlir.
I think just like LLVM, MLIR will have frequent api changes across releases, so maintenance like this PR will be needed.

---

To refresh, this PR now covers (tested on ) llvm/mlir: 15.0.7, 16.0.6 and 17.0.0rc3.

[junrushao]

Thanks @cbalint13! This is merged!