[TIR] Fixup max/min with input containing NAN

[actinks]

llvm.minimum and llvm.maximum: Return NaN if one the arguments is NaN.
refs: https://llvm.org/docs/LangRef.html#floating-point-min-max-intrinsics-comparison

fixed #19579

[gemini-code-assist]

Code Review

This pull request updates the LLVM code generation for minimum and maximum operations to use LLVM's CreateMinimum and CreateMaximum instructions for non-integer types. Feedback highlights that these instructions require LLVM 12 or later, so version guards are necessary to prevent compilation failures on older supported versions. It is also suggested to restrict these calls specifically to floating-point types for better type safety.

[actinks]

@tqchen The failing test case in ir_module.py is due to the LLVM backend lagging behind LLVM IR in its support for llvm.{min,max}imum.{f32,f64}. How should we handle this situation?
Related issue: llvm/llvm-project#53353

LLVM ERROR: Cannot select: 0x5d50add04bf0: f32 = fmaximum 0x5d50add04d28, ConstantFP:f32<0.000000e+00>, IRModule.CodeGenLLVM:0

The fmaximum instruction is supported in llvm 18.1.0 for x86-64 and in llvm 19.1.0 for i686.
proof:
X86-64: https://godbolt.org/z/b95xvbKxn
i686: https://godbolt.org/z/q39dr1bs9

[cbalint13]

@tqchen The failing test case in ir_module.py is due to the LLVM backend lagging behind LLVM IR in its support for llvm.{min,max}imum.{f32,f64}. How should we handle this situation? Related issue: llvm/llvm-project#53353

LLVM ERROR: Cannot select: 0x5d50add04bf0: f32 = fmaximum 0x5d50add04d28, ConstantFP:f32<0.000000e+00>, IRModule.CodeGenLLVM:0

The fmaximum instruction is supported in llvm 18.1.0 for x86-64 and in llvm 19.1.0 for i686. proof: X86-64: https://godbolt.org/z/b95xvbKxn i686: https://godbolt.org/z/q39dr1bs9

Hi @actinks !

• I dont think it is a good idea to enfoce everything via createmin/createmax at IR / LLVM-IR level.
• We also loose backward compatibility with older LLVM (I personally highly appreciate this one in TVM).

---

Instead we could do:

• Add new TIRx functions nanmin, nanmax just like the existing nearbyint see https://github.com/apache/tvm/blob/main/src/target/llvm/intrin_rule_llvm.cc#L95
• These new nanmin and nanmax can offer TVM / TIRx to be on-par with numpy functions : https://numpy.org/devdocs/reference/generated/numpy.nanmin.html

Now, these new functions can be used to "protect" needed things in the higher level graph flow.
These new TIRx can be available only for certain LLVM version above (protecting with #if LLVM_VERSION_MAJOR).

@tqchen what do you think ?

---

Later EDIt:

These new TIRx can be available only for certain LLVM version above (protecting with #if LLVM_VERSION_MAJOR).

I think with workaround (emmiting multiple llvm insn) it can be made fully backward to any LLVM version.
But not so sure how to replicate nanmin / nanmax to opencl, cuda, vulkan and all backend beyond llvm.