mirror of
https://github.com/tinygrad/tinygrad.git
synced 2026-08-29 12:56:07 +00:00
@@ -822,6 +822,7 @@ class TestOps(unittest.TestCase):
|
||||
helper_test_op([], lambda: tor0&tor1, lambda: ten0&ten1, forward_only=True)
|
||||
|
||||
helper_test_op(None, lambda x: (1 < x) & (x < 2), forward_only=True, vals=[[1.2, 1.2, 1.2, 3.2]])
|
||||
helper_test_op([(3000,)]*10, lambda *xs: (sum(xs[1:], xs[0]) > 5) & (xs[0] < 0.9), forward_only=True)
|
||||
|
||||
if not COMPILE_ONLY:
|
||||
np.testing.assert_equal((Tensor(2**64-1, dtype=dtypes.uint64) & 0xFFFFFFFF).numpy(), 0xFFFFFFFF)
|
||||
|
||||
@@ -637,10 +637,12 @@ def encode(x:UOp, opc:int, reg:int|None=None, pp:int=0, sel:int=0, we:int=0) ->
|
||||
if sz == 2: inst += bytes([0x66])
|
||||
# bit signaling 64 bit variant of instruction
|
||||
w = sz == 8
|
||||
# REX byte is required when 64 bit or an extended reg is used (index 8 - 15) or lower 8 bits of (rsp, rbp, rsi, rdi) are accessed
|
||||
if w | r | _x | b | (reg_sz == 1 & reg >> 2) | (rm_sz == 1 & rm >> 2): inst += bytes([0b0100 << 4 | w << 3 | r << 2 | _x << 1 | b])
|
||||
# legacy 8bit opcode is 1 less than 16-64bit variants
|
||||
if (rm_sz == 1 or reg_sz == 1) and x.arg not in X86GroupOp.ReadFlags | {X86Ops.LEA}: opc -= 1
|
||||
demote = (rm_sz == 1 or reg_sz == 1) and x.arg not in X86GroupOp.ReadFlags | {X86Ops.LEA}
|
||||
# REX byte is required when 64 bit or an extended reg is used (index 8 - 15) or lower 8 bits of (rsp, rbp, rsi, rdi) are accessed
|
||||
if w | r | _x | b | (reg_sz == 1 & reg >> 2) | (rm_sz == 1 & rm >> 2) | (demote and disp_uop is None and rm >= 4):
|
||||
inst += bytes([0b0100 << 4 | w << 3 | r << 2 | _x << 1 | b])
|
||||
if demote: opc -= 1
|
||||
# OPCODE byte
|
||||
inst += opc.to_bytes((opc.bit_length() + 7) // 8, 'big')
|
||||
# MODRM byte
|
||||
|
||||
Reference in New Issue
Block a user