fldpi only loads pi to 53-bit accuracy instead of 64-bit accuracy

[pcdosretro]

Code of Conduct & Contributing Guidelines

• I agree to follow the code of conduct and the contributing guidelines.

Have you checked that no other similar bug report(s) already exists?

• I have searched and didn't find any similar issues.

What operating system(s) this bug have occurred on?

Windows 7 SP 1 (64-bit)

What version(s) of DOSBox-X have this bug?

0.83.25, 0.84.0 (64-bit builds only)

Describe the bug

fldpi only loads pi to 53-bit accuracy instead of 64-bit accuracy. The fraction loaded is C90FDAA2 2168C000

This is also true of the instructions fldl2t, fldl2e, fldlg2, fldln2

Expected behavior

fldpi should load the fraction C90FDAA2 2168C234

Steps to reproduce the behaviour

Get DEBUG or DEBUGX from my website - https://pcdosretro.github.io/enhdebug.htm

a
fninit
fldpi
fstp tbyte ptr [80]
int 3

g
d 80 la

Used configuration

No response

Emulator log

No response

Additional context

No response

[cimarronm]

On anything compiling for an x86 platform with the C_FPU_X86 define set, it will use the fldpi instruction itself in emulation so this shouldn't be a problem. On compiler/platform combinations which support a long double type, PR #3453 should allow it to fill the full 64-bit mantissa. BTW, I am getting a mantissa of 0xc90fdaa22168c235 (instead of 0xc90fdaa22168c234 as you listed) which is closer to PI.

0x4000c90fdaa22168c235 = 3.14159265358979323851...
0x4000c90fdaa22168c234 = 3.14159265358979323829...
PI = 3.14159265358979323846...

[pcdosretro]

The Intel manual says FLDPI loads C90FDAA2 2168C234 C which gives you 3.141592653589793238
I will recheck it with v0.83.26 which should have your fix for this (and hopefully a fix for my previous FPU bug #3440)

[pcdosretro]

Not fixed in DOSBox-X 0.84.0 - somewhere long double values are getting demoted to double.

[pcdosretro]

The problem here is the damn MSVC compilers don't actually support long double as 80-bit floating point, long double is the same as double. The sad thing is the old 16-bit DOS MS C compilers did in fact support long double as 80-bit floating point. Another problem is that MSVC doesn't support 64-bit inline assembly so the inline assembly in fpu_instructions_x86.h can't be used.

The 32-bit builds of DOSBox-X work fine.

[cimarronm]

Yeah, unfortunately microsoft decided not to support the extended precision long double nor inline assembly in the 64-bit version of the compiler. The MinGW 64-bit version for windows should support C_FPU_X86 and give you a 64-bit version with the x87 instruction set.

[pcdosretro]

I switched to the 32-bit build which works fine. Technically this isn't a bug in DOSBox-X so I guess this can be closed.