forked from M-Labs/nac3
1
0
Fork 0
nac3/nac3standalone/demo/src/min_max_abs.py

92 lines
2.3 KiB
Python

@extern
def output_int32(x: int32):
...
@extern
def output_uint32(x: uint32):
...
@extern
def output_int64(x: int64):
...
@extern
def output_uint64(x: uint64):
...
@extern
def output_float64(x: float):
...
def run() -> int32:
# min ===========
output_int32(int32(min(True, False)))
output_int32(int32(min(True, True)))
output_int32(int32(min(False, False)))
output_int32(min(-12, 0))
output_int32(min(1, -3))
output_int32(min(2, 2))
output_int64(min(int64(-12), int64(0)))
output_int64(min(int64(1), int64(-3)))
output_int64(min(int64(2), int64(2)))
output_uint32(min(uint32(12), uint32(0)))
output_uint32(min(uint32(12), uint32(24)))
output_uint32(min(uint32(2), uint32(2)))
output_uint64(min(uint64(12), uint64(0)))
output_uint64(min(uint64(12), uint64(24)))
output_uint64(min(uint64(2), uint64(2)))
output_float64(min(-12.234, 3.23))
output_float64(min(0.1, 12.3))
output_float64(min(1.1, 1.1))
# max ===========
output_int32(int32(max(True, False)))
output_int32(int32(max(True, True)))
output_int32(int32(max(False, False)))
output_int32(max(-12, 0))
output_int32(max(1, -3))
output_int32(max(2, 2))
output_int64(max(int64(-12), int64(0)))
output_int64(max(int64(1), int64(-3)))
output_int64(max(int64(2), int64(2)))
output_uint32(max(uint32(12), uint32(0)))
output_uint32(max(uint32(12), uint32(24)))
output_uint32(max(uint32(2), uint32(2)))
output_uint64(max(uint64(12), uint64(0)))
output_uint64(max(uint64(12), uint64(24)))
output_uint64(max(uint64(2), uint64(2)))
output_float64(max(-12.234, 3.23))
output_float64(max(0.1, 12.3))
output_float64(max(1.1, 1.1))
# abs ===========
output_int32(int32(abs(False)))
output_int32(int32(abs(True)))
output_int32(abs(0))
output_int32(abs(-3))
output_int32(abs(2))
output_int64(abs(int64(-12)))
output_int64(abs(int64(0)))
output_int64(abs(int64(2)))
output_uint32(abs(uint32(0)))
output_uint32(abs(uint32(24)))
output_uint64(abs(uint64(0)))
output_uint64(abs(uint64(24)))
output_float64(abs(-12.234))
output_float64(abs(-0.0))
output_float64(abs(1.1))
return 0