Re: Phix: peek8u deprecated.
- Posted by petelomax Jan 01, 2017
- 1532 views
Oh, ffs - I read something on the internet and it was WRONG! 
I knew there was a problem with unsigned 64 bit ints in 80-bit floats, so just accepted his word (/diagram) as the gospel truth.
(Figure 2.3 shows a literal 1, which I thought was odd, but y'know the whole "well, that explains it" thing.. It ain't a fixed 1.)
To my shame, I already knew there was a problem getting them in, what I missed was the (simple) problem in getting them out.
The FP instructions fild/fist/fistp work with *integers* there is no instruction for unsigned integers.
The Floating Point Unit doesn't know of unsigned integers existence, so if the number won't fit in a signed integer then it is an overflow (0x80000000....).
There may be a better way (suggestions welcome) but this will load an unsigned int (already in rax):
-- to load unsigned, right shift rax by 1, save odd bit in rcx, then *2+[0|1]
xor rcx,rcx
shr rax,1
rcl rcx,1
push rax
push rcx
fild qword[rsp]
fild qword[rsp+8]
add rsp,16
fadd st0,st0 -- *2
faddp -- +0|1
Extracting an unsigned int (from an 80-bit float already in st0) is not really any better or worse:
-- if uint>#7FFF... then uint-=#1_0000...
push r15 -- #4000_0000_0000_0000
fild qword[rsp]
fadd st0,st0 -- #8000_0000_0000_0000
fld st1
fcomp
fnstsw ax
sahf
jb @f
fadd st0,st0 -- #1_0000_0000_0000_0000
fsub st1,st0
@@:
fstp st0 -- discard
fistp qword[rsp]
pop rax
(Obviously, do a mov r15,h4 first when unsure, but that will be unneccessary more often than not.)
[Some further care may also be needed when dealing with already signed values, and it may be easier in hll]
So that just needs applying in ten dozen or so places throughout the backend, and peek8u can live on 
Pete
