Scribd
Upload
49 views

Extended Precision FPUDemo

This document demonstrates how code was implemented in the FPUExtendedCalculator to improve the floating point precision of Windows from 15 to up to 20 digits. It shows the buttons that are used to input constants or real numbers and store them in FPU registers. It explains that pressing buttons in the correct order allows observing changes in FPU registers and memory with the new code. Instructions are provided in the CodeChanges.txt file and more details can be found at the given GitHub URL.

Uploaded by

jose carlos
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as RTF, PDF, TXT or read online on Scribd
49 views

Extended Precision FPUDemo

This document demonstrates how code was implemented in the FPUExtendedCalculator to improve the floating point precision of Windows from 15 to up to 20 digits. It shows the buttons that are used to input constants or real numbers and store them in FPU registers. It explains that pressing buttons in the correct order allows observing changes in FPU registers and memory with the new code. Instructions are provided in the CodeChanges.txt file and more details can be found at the given GitHub URL.

Uploaded by

jose carlos
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as RTF, PDF, TXT or read online on Scribd
You are on page 1/ 3

How to use

This program serves as a demo to demonstrate how the new code implemented in the
FPUExtendedCalculator can improve the original 15-digit significant precision of the
Windows system to up to 20 digits.

There is a button for each constant implemented in the FPU code and another one to
enter any real number.

Once a constant button is pressed (or a real number is entered), the value is transformed
into an integer with at least 15 significant digits in the integer part and is stored in the FPU
st(0) register as the original code do.

Next, to see the new code in action, the button 1O with the FPU code must be pressed,
and so on. The changes in the FPU 16-bit registers, the FPU 80-bit registers, and in
memory (if applicable) can be observed.

It is essential to follow the correct order. Otherwise, strange results may be obtained.
However, you can reinitialize the process by pressing a new constant button or the real
number button.

The new code to compile the FPUExtendedCalculator is also available in the file
CodeChanges.txt.

(*) In the original code, the real number that had to be shown, had been
previously rounded and then transform in a integer with at least 15 significant
digits (doubled precision), and it is into the st0 FPU register just ready to be
store in memory in 18 digits BCD (Binary Coded Decimal) format. With the new, if
the flag of precision is not set, we avoid the initial rounded code and then obtain
the decimal part and store it in memory just before do the same with the integer
part to get an extra precision of the real number. Then we avoid again other
rounded made in the original code (in normal asm). Indeed the memory space
used for that is used now to insert the new code.

// Previous frndint (now only if selected in program menu)


00409C0B | EB 79 | jmp 409C86
//Next code WITHOUT CHANGE
00409C0D | 8B7D FC | mov edi,dword ptr ss:[ebp-4]
00409C10 | DBAF 54B84700 | fld st(0),tword ptr ds:[edi+47B854]
00409C16 | D8D9 | fcomp st(0),st(1)
00409C18 | 9B | fwait
00409C19 | DD7D F6 | fnstsw word ptr ss:[ebp-A]
00409C1C | 9B | fwait
00409C1D | 66:F745 F6 0041 | test word ptr ss:[ebp-A],4100
00409C23 | 74 09 | je 409C2E
00409C25 | DAB7 60B84700 | fidiv st(0),dword ptr ds:[edi+47B860]
00409C2B | FF45 F8 | inc dword ptr ss:[ebp-8]
// (*)
00409C2E | 8D7B 03 | lea edi,dword ptr ds:[ebx+3]
//value store in st0 store also in st(1) and st(2)
00409C31 | D9C0 | fld st(0),st(0)
00409C33 | D9C0 | fld st(0),st(0)
00409C35 | 9B | fwait
//save FPU register control word, in memory
00409C36 | D97D E6 | fnstcw word ptr ss:[ebp-1A]
//set FPU control word RC bits to (11) -> truncate
00409C39 | 66:814D E6 000C | or word ptr ss:[ebp-1A],C00
//load FPU register control word with the new set
00409C3F | D96D E6 | fldcw word ptr ss:[ebp-1A]
//with new setting frndint truncate st0 register value
00409C42 | D9FC | frndint
//subtract st1 minus st0 and store in st0 (decimal part)
00409C44 | DEE9 | fsubp st(1),st(0)
//store $64 (decimal 100) in ax
00409C46 | 66:B8 6400 | mov ax,64
//move ax to memory
00409C 4A | 66:8945 E6 | mov word ptr ss:[ebp-1A],ax
//multiply value in st0 by decimal 100
00409C4E | DE4D E6 | fimul st(0),word ptr ss:[ebp-1A]
//store decimal part of st0 value into 18 digits BCD
00409C51 | DF75 E8 | fbstp tword ptr ss:[ebp-18],st(0)
//store integer part of st0 value into 18 digits BCD
00409C54 | DF75 E9 | fbstp tword ptr ss:[ebp-17],st(0)
//Check pending unmasked floating-point exception
00409C57 | 9B | fwait
//Initialize FPU (very important!)
00409C58 | DBE3 | fninit
//now 12 pairs of figures (BDC 10+2)
00409C5A | BA 0A000000 | mov edx,A
// WITHOUT CHANGE
00409C5F | 8A442A E7 | mov al,byte ptr ds:[edx+ebp-19]
00409C63 | 88C4 | mov ah,al
00409C65 | C0E8 04 | shr al,4
00409C68 | 80E4 0F | and ah,F
00409C6B | 66:05 3030 | add ax,3030
00409C6F | 66:AB | stosw
00409C71 | 4A | dec edx
00409C72 | 75 EB | jne 409C5F
00409C74 | 32C0 | xor al,al
00409C76 | AA | stosb
00409C77 | 8B7D F8 | mov edi,dword ptr ss:[ebp-8]
00409C7A | 037D 08 | add edi,dword ptr ss:[ebp+8]
00409C7D | 79 16 | jns 409C95
00409C7F | 31C0 | xor eax,eax
00409C81 | E9 52FFFFFF | jmp 409BD8
//is the flag set?
00409C86 | 3A05 B8D14700 | cmp al,byte ptr ds:[47D1B8]
//if set avoid round
00409C8C | 75 02 | jne 409C90
//if not set round
00409C8E | D9FC | frndint
00409C90 | E9 78FFFFFF | jmp 409C0D
//Now 19 figures plus (,)
00409C95 | BF 14000000 | mov edi,14

The Delphi project code, and extensive explanation can be found at:

https://github.com/GutiSoftware/Extended-precision-in-Delphi

You can mail me for any doubt in: [email protected]

GutiSoft
March 2023

You might also like