arch: x86: add translation from roundsd to IR

This change adds three new ops for rounding:
- 'fpround_towardsnearest': rounds to nearest int, even for ties
- 'fpround_down': rounds to nearest int <= float
- 'fpround_up': rounds to nearest int >= float
- 'fpround_towardszero'(existing): rounds to nearest int s.t. |int| <=
                                   |float|

For a variant of roundsd that uses mxcsr register as the rounding mode,
currently we assume it's fpround_towardsnearest. This may cause trouble
if the program modifies MXCSR register.

Tested:
$ objdump -dSC rounds.so
0000000000001000 <entry>:
    1000:	66 0f 3a 0b c0 01    	roundsd $0x1,%xmm0,%xmm0
    1006:	66 0f 3a 0a c0 0c    	roundss $0xc,%xmm0,%xmm0
    100c:	c3                   	retq

$ cat test.py
from miasm.analysis.binary import *
from miasm.analysis.machine import Machine
from miasm.expression.simplifications import expr_simp
with open('rounds.so', 'rb') as f:
    cont = ContainerELF.from_stream(f)
machine = Machine(cont.arch)
mdis = machine.dis_engine(cont.bin_stream, loc_db=cont.loc_db)
asmcfg = mdis.dis_multiblock(cont.loc_db.get_name_offset('entry'))
ir = machine.ir(asmcfg.loc_db)
ircfg = ir.new_ircfg_from_asmcfg(asmcfg)
ircfg.simplify(expr_simp)
print(ircfg.blocks[ircfg.heads()[0]].to_string())

$ python test.py
loc_key_0:
XMM0 = {fpround_down(XMM0[0:64]) 0 64, XMM0[64:128] 64 128}
XMM0 = {fpround_towardsnearest(XMM0[0:32]) 0 32, XMM0[32:128] 32 128}
RSP = RSP + 0x8
RIP = @64[RSP]
IRDst = @64[RSP]

0 files changed, 0 insertions, 0 deletions