reko
Looking at the output one can see:- high: reko doesn’t load constant values (neither float nor int) from read-only sections
- reko doesn’t evaluate functions like pow or fabs with constant arguments
- reko doesn’t perform basic arithmetic operations like 1.0+1.0
- fbstp is decompiled as a cast but the operation is more complex
- The output of FCMOV (
if (Test(LT,SLICE(SLICE(dwArg04, byte, 8), bool, 1)))) is a bit too magic for me - fcmovu for unordered is unfinished (seeing only P). Also fcmovnu is the same as fcmovu
- looks like fcmovnb is decompiled as fcmovb
- fcom uses a magic cond() function
- additional noise in integer code in the fcom function
- output for fcomi and ftst is incomplete
- fist and frndint doesn’t respect rounding mode
- wrong parameter values for fpatan
- looping because of partial remainder looks wrong
- leaves unused calls for fptan, fsincos, fxtract
- wrong parameter used for fyl2x
- modelling of the status word is incorrect/not easy to understand
retdec
- wrong modelling of f2xm1
- doesn’t evaluate functions like pow or fabs with constant arguments
- confusing use of float80_t for fbst
- wrong decompilation of fbld
- missing code for fcom and ftst
- gets confused in fcomi, fptan and fsincos
- fist and frndint doesn’t respect rounding mode
- fld for 32bit floats loads wrong number
- bad modeling of fpatan with only one parameter
- incorrect implementation of fprem1
- doesn’t model C2 change of fprem
- missing memory write for 32bit fst
- missing slot is inf for fxam
- multiple copies of __pseudo functions for fxtract
- superfluous calls in fxtract
- modelling of the status word is incorrect/not easy to understand
ghidra
- displays floating point arguments to printf as hex
- fbst is a simple assignment
- fbld is incomplete
- fcom and fcomi doesn’t test for unordered
- doesn’t evaluate functions like fcos with constant arguments
- fist and frndint doesn’t respect rounding mode
- fld a 80bit float doesn’t work
- fprem and fprem1 not working
- doesn’t model that fprem updates the status word
- fxtract,fyl2x, fyl2xp1 is just identity
- modelling of the status word is incorrect/not easy to understand
Summary
All examined decompilers model the x87 instructions to a certain degree which is probably good enough for common binaries. There are however in all decompilers x87 details which are not supported: rounding modes, handling of unordered compares and the BCD format.
Thank you for reading and please send questions or feedback via email to holdec@kronotai.com or contact me on Twitter.
]]>