Dynamic core FPU emulation/JIT broken on ARM64/Apple Silicon macOS

[badlogic]

Describe the bug

The dynamic core FPU JIT/emulation appears to be broken on Apple Silicon when using DOSBox-x for Apple Silicon. This trivial C program

#include <stdio.h>

int main(void) {
    double v = 123.0;
    printf("%f", v);
    return 0;
}

will "print" the following

Steps to reproduce the behaviour

1. Compile the above program with DJGPP (any version will do, I tested 12.1.0 and 7.5.0). I've attached an executable: dos.zip
2. Run the executable using DOSBox-X for Apple Silicon on an Apple Silicon macOS machine
3. See

Expected behavior

The program should print 123.0.

Note that DOSBox-x x86_64 for macOS running on Apple Silicon via Rosetta 2 works as intended.

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

macOS on Apple Silicon

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

0.84.1, latest commit on GitHub

Used configuration

No response

Output log

LOG: Early LOG Init complete
LOG: "65536" is not a valid value for variable: pcrate.
It might now be reset to the default value: 65536
LOG: DOSBox-X's working directory: /Users/badlogic/workspaces/minifb
LOG: Logging init: beginning logging proper. This is the end of the early init logging
LOG: Logging: No logfile was given. All further logging will be discarded.
LOG: DOSBox-X version 0.84.1 (macOS SDL2)
LOG: macOS EXE path: dosbox-x
LOG: macOS Resource path: 
LOG: macOS: NSTouchBar support detected in system
LOG: Host keyboard layout is now us (US English)
LOG: Mapper keyboard layout is now us (US English)
LOG: SDL2 reports desktop display mode 2056 x 1329
LOG: The default output for the video system: surface
LOG: Configured windowposition: -
LOG: SDL: Current window pixel format: SDL_PIXELFORMAT_ARGB8888
LOG: Screen report: Method 'CoreGraphics' (2056.000 x 1329.000 pixels) at (0.000 x 0.000) (344.702 x 222.816 mm) (13.571 x 8.772 in) (151.500 x 151.500 DPI)
LOG: ISA BCLK: 8333333.333Hz (25000000/3)
LOG: monopal: green, 
LOG: Active save slot: 1 [Empty]
LOG: USING AVI+ZMBV
LOG: Max 3670016 sz 16384
LOG: Final 16384
LOG: MIDI:Opened device:coreaudio
LOG: Pentium CMPXCHG8B emulation is enabled
LOG: VOODOO LFB now at d0000000
LOG: Serial1: BASE 3f8h
LOG: Serial2: BASE 2f8h
LOG: Parallel1: BASE 378h
LOG: MPU-401 Registering I/O ports as if IBM PC MPU-401 at base 330h
LOG: Screen report: Method 'CoreGraphics' (2056.000 x 1329.000 pixels) at (0.000 x 0.000) (344.702 x 222.816 mm) (13.571 x 8.772 in) (151.500 x 151.500 DPI)
LOG: Allocated APM BIOS pm entry point at f000:ce20
LOG: Writing code to fce20
LOG: Writing code to fce40
LOG: ETHERNET: NE2000 Ethernet emulation backend selected: slirp
LOG: SLIRP: Slirp version: 4.7.0
LOG: SLIRP: Successfully initialized
LOG: NE2000: Base=0x300 irq=3
LOG: Screen report: Method 'CoreGraphics' (2056.000 x 1329.000 pixels) at (0.000 x 0.000) (344.702 x 222.816 mm) (13.571 x 8.772 in) (151.500 x 151.500 DPI)
LOG: ISA Plug & Play BIOS enabled
LOG: VGA ROM BIOS init callback
LOG: WARNING: No translation support (to host) for code page 0
LOG:          0 ERROR BIOS:Keyboard layout file auto not found
LOG:          0 ERROR BIOS:Keyboard layout file auto not found
LOG: XMS: 50 handles allocated for use by the DOS environment
LOG: EMS page frame at 0xe000-0xefff
LOG: COMMAND.COM env size:             720 bytes
LOG: COMMAND.COM environment block:    0x0701 sz=0x002d
LOG: COMMAND.COM main body (PSP):      0x072f sz=0x009a
LOG: COMMAND.COM stack:                0x0749
LOG: pixratio 1.350, dw false, dh false
LOG: Aspect ratio: 720 x 540  xToY=1.333 yToX=0.750
LOG: surface consider=720x400 final=720x400
LOG: Screen report: Method 'CoreGraphics' (2056.000 x 1329.000 pixels) at (0.000 x 0.000) (344.702 x 222.816 mm) (13.571 x 8.772 in) (151.500 x 151.500 DPI)
LOG: SHELL:Redirect output to nul
LOG: DOSBox-X has switched to max cycles, because of the setting: cycles=auto.
If the game runs too fast, try a fixed cycles amount in DOSBox-X's options.

### Additional information

_No response_

### Have you checked that no similar bug report(s) exist?

- [X] I have searched and didn't find any similar bug report.

### Code of Conduct & Contributing Guidelines

- [X] I agree to follow the code of conduct and the contributing guidelines.

[emendelson]

I can confirm this with current code, in both the SDL1 and SDL2 builds. I get the correct result in an Apple Silicon build I made on 1 April 2022, but the wrong result in an Apple Silicon build I made on 3 September 2022. I can't narrow it down beyond that, but the regression clearly happened in April through July of this year.

[mariobrx]

Did you try to put fpu=false in the CPU section of conf file ? I also had calculation issues on RaspiOS 64bit in VisualBasic for DOS (see https://github.com/joncampbell123/dosbox-x/issues/3872)

[emendelson]

Yes, fpu=false has no effect on this bug.

[joncampbell123]

If you can compile from source, try the latest commit. Some floating point fixes have been made.