remove "no-nans-fp-math"="true" for LLVM (#5282)

fixed isnan for llvm (still have issue with < nan)

test/external/external_test_onnx_backend.py

@@ -171,17 +171,6 @@ if Device.DEFAULT == "METAL" or (OSX and Device.DEFAULT == "GPU"):
   backend_test.exclude('test_mish_cpu')
   backend_test.exclude('test_mish_expanded_cpu')
 
-# TODO: llvm has problems with inf
-if Device.DEFAULT in ['LLVM']:
-  backend_test.exclude('test_isinf_cpu')
-  backend_test.exclude('test_isinf_negative_cpu')
-  backend_test.exclude('test_isinf_positive_cpu')
-
-# # TODO: problems with nan
-if Device.DEFAULT in ['LLVM']:
-  backend_test.exclude('test_isnan_float16_cpu')
-  backend_test.exclude('test_isnan_cpu')
-
 # disable model tests for now since they are slow
 if not getenv("MODELTESTS"):
   for x in backend_test.test_suite:

test/test_dtype_alu.py

@@ -24,7 +24,6 @@ binary_operations = [operator.add, operator.sub, operator.mul, operator.lt, oper
 # TODO: LLVM comparing with nan is incorrect
 if Device.DEFAULT == "LLVM":
   binary_operations.remove(operator.lt)
-  binary_operations.remove(operator.eq)
 
 integer_binary_operations = binary_operations + [(Tensor.xor, np.bitwise_xor), (Tensor.bitwise_and, np.bitwise_and),
                                                  (Tensor.bitwise_or, np.bitwise_or)]

tinygrad/renderer/llvmir.py

@@ -86,10 +86,6 @@ class LLVMRenderer(Renderer):
     for a in func.args:
       if a.type.is_pointer: a.add_attribute("noalias")
 
-    # add the function attribute "no-nans-fp-math"="true", which informs llvm that it allowed to use vectorization optimizations
-    func.attributes._known = func.attributes._known.union(frozenset(['"no-nans-fp-math"="true"']))
-    func.attributes.add('"no-nans-fp-math"="true"')
-
     bb = [ir.IRBuilder(func.append_basic_block("entry"))]
     loop_blocks: List = []
     reduce_phis: List = []