def __init__(self):
# TODO: come up with a less naive way of handling the symtab and types
self.classes = None
self.symtab = {}
self.typetab = {}
self.is_break = False
self.current_class = None
self.loop_end_blocks = []
self.loop_cond_blocks = []
context = ir.Context()
self.module = Module(name='opal-lang', context=context)
self.blocks = []
self.scope = {}
self._add_builtins()
func_ty = ir.FunctionType(ir.VoidType(), [])
func = Function(self.module, func_ty, 'main')
self.current_function = func
entry_block = self.add_block('entry')
exit_block = self.add_block('exit')
self.function_stack = [func]
self.builder = Builder(entry_block)
self.exit_blocks = [exit_block]
self.block_stack = [entry_block]
def test_nvvm_from_llvm(self):
m = Module("test_nvvm_from_llvm")
fty = Type.function(Type.void(), [Type.int()])
kernel = m.add_function(fty, name='mycudakernel')
bldr = Builder(kernel.append_basic_block('entry'))
bldr.ret_void()
set_cuda_kernel(kernel)
fix_data_layout(m)
ptx = llvm_to_ptx(str(m)).decode('utf8')
self.assertTrue('mycudakernel' in ptx)
if is64bit:
self.assertTrue('.address_size 64' in ptx)
else:
self.assertTrue('.address_size 32' in ptx)
def test_inline_rsqrt(self):
mod = Module(__name__)
fnty = Type.function(Type.void(), [Type.pointer(Type.float())])
fn = mod.add_function(fnty, 'cu_rsqrt')
bldr = Builder(fn.append_basic_block('entry'))
rsqrt_approx_fnty = Type.function(Type.float(), [Type.float()])
inlineasm = InlineAsm.get(rsqrt_approx_fnty,
'rsqrt.approx.f32 $0, $1;',
'=f,f',
side_effect=True)
val = bldr.load(fn.args[0])
res = bldr.call(inlineasm, [val])
bldr.store(res, fn.args[0])
bldr.ret_void()
# generate ptx
nvvm.fix_data_layout(mod)
nvvm.set_cuda_kernel(fn)
nvvmir = str(mod)
ptx = nvvm.llvm_to_ptx(nvvmir)
self.assertTrue('rsqrt.approx.f32' in str(ptx))