Starting p-code interpreter seems to take 1 or 2 mins on DJGPP

[dufresnep]

Running on FreeDOS 1.3 using DJGPP.

With gnat 10.30, that does not support "gnatmake -P hac" I did (added -m 300 to not miss QEMU memory):
gnatmake src/apps/hac.adb -Isrc -Isrc/compile -I src/compile/emit -Isrc/manage -Isrc/execute

I got a hac.exe 3,171,371 bytes.

CD EXM
../hac -v2 hello.adb
compiles in .11 seconds... but take a bit more than 1 minute I think on:
Starting p-code VM interpreter....

before writing Hello world!
seems to block there for about 1 or 2 mins before showing Execution of hello.adb completed.

Similar with ../hac -v2 series.adb
compiles fast, take 1 or 2 mins before executing... then run "normal" fast.

[dufresnep]

The AI have finally been able to use me to make a MINGW MSYS2 GCC/gnat 12.2.0 cross-compiler to DJGPP and build an hac_dos.exe file (that I did not tested yet).

On HAC side, I had to add a file:
Utilisateur@DESKTOP-TPTM1G7 MINGW64 ~/build-gcc/hac
$ cat math_stubs.c

#include <math.h>

/* Redirection des fonctions Long Double vers Double */
long double sinl(long double x)  { return (long double)sin((double)x); }
long double cosl(long double x)  { return (long double)cos((double)x); }
long double logl(long double x)  { return (long double)log((double)x); }
long double expl(long double x)  { return (long double)exp((double)x); }
long double atanl(long double x) { return (long double)atan((double)x); }
long double powl(long double x, long double y) { return (long double)pow((double)x, (double)y); }
long double tanl(long double x)  { return (long double)tan((double)x); }
long double asinl(long double x) { return (long double)asin((double)x); }
long double acosl(long double x) { return (long double)acos((double)x); }
long double tanhl(long double x) { return (long double)tanh((double)x); }
long double sinhl(long double x) { return (long double)sinh((double)x); }
long double coshl(long double x) { return (long double)cosh((double)x); }

And use it to build hac_dos.exe with:
$ i586-pc-msdosdjgpp-gnatmake -f -a -O2 src/apps/hac.adb -Isrc -Isrc/compile -Isrc/compile/emit -Isrc/manage -Isrc/execute -o hac_dos.exe -largs math_stubs.o -lm

On the build of the cross-compiler, it was quite hard... had to make a lot of dummy functions for poll.c and tasking functions.

Will soon test the new binary to see it it is slow too to start p-code interpreter.

[dufresnep]

Oh... In VirtualBox both versions are reasonably fast, both versions executing in about 5.5 seconds for hello.adb... means I spend so much time doing a cross-compiler for nothing!

[dufresnep]

I just tested my hac_dos.exe (from https://gitlab.com/dufresnep/hac ) in QEMU on Ubuntu 25.10.
hac_dos.exe -v2 HELLO.ADB compiled in 0.11 secs, and executed in 1149 seconds (19 mins).

paul@starman:~/hac_project$ qemu-system-x86_64 -enable-kvm -m 64 -hda myhac.img -cdrom hac_directory.iso
WARNING: Image format was not specified for 'myhac.img' and probing guessed raw.
Automatically detecting the format is dangerous for raw images, write operations on block 0 will be restricted.
Specify the 'raw' format explicitly to remove the restrictions.
qemu-system-x86_64: warning: host doesn't support requested feature: CPUID[eax=80000001h].ECX.svm [bit 2]

When I remove the -enable-kvm, ..\HAC_DOS.EXE HELLO.ADB returns
Debug at eip=fffffb72: flags=3246.

[zertovitch]

This delay is strange indeed.
Perhaps the initialization of the memory of the p-code virtual machine takes for mysterious reasons much more time on your DOS setup. A discriminant is initialized in each memory cell of that VM (there are 5 million of them), and my guess is that may be the cause.

Last time I tested GNAT on DOS (DJGPP) must have been in ~2016, for testing Engine_3D.
Previously, I had used extensively GNAT for DOS till the early 2000's.
I used the following combinations OS/DPMI:

• DR-DOS single-task & CWSDPMI
• DR-DOS multi-task & and its own DPMI implementation, but IIRC it crashed quite quickly...
• MS-DOS & CWSDPMI
• Windows NT's NTVDM and its own DPMI implementation.