gcc 3.2.3 i370 with z/PDOS-generic and mfemul

[Paul Edwards via Gcc]

Hello.

I have a mainframe operating system called z/PDOS-generic
available from https://pdos.org

I have ported a slightly modified gcc 3.2.3 (called
gccmvs) to it, and when run under Hercules/380
(mainframe emulator) it works fine and gccmvs is
able to reproduce itself byte-exact.

But I have a new emulator called mfemul.c, and it
isn't very mature and almost certainly has a bug in
it that is affecting gcc 3.2.3. And since that is the only
known issue at this stage, this is where I am trying
to debug, which means delving into the internals
of gcc to some degree.

Since I have a good (both PC and Hercules/380)
version and a bad version, I can put in printf to
compare what is happening, without actually having
to understand the internals.

However, I'm not sure where to get started on this one.

I do know the "-da" and that has shown that between:

tempr.c.22.postreload
and
tempr.c.23.flow2

this "plus" has gone away:

(insn 84 15 16 (set (reg/v:SI 3 3 [27])
        (mem/f:SI (plus:SI (reg/f:SI 11 11)
                (const_int 4 [0x4])) [3 ymode+0 S4 A32])) 15 {movsi} (nil)
    (expr_list:REG_EQUIV (mem/f:SI (plus:SI (reg/f:SI 11 11)
                (const_int 4 [0x4])) [3 ymode+0 S4 A32])
        (nil)))

(it's still there on the PC)

This is the code I am compiling:

D:\devel\gcc\gcc>type tempr.c
extern unsigned char mode_size[50];

int foo(int xmode, int ymode)
{
  int mode_multiple;

  mode_multiple = mode_size[xmode] / mode_size[ymode];
  if (mode_multiple == 0)
  {
    for (;;) ;
  }
  return 0;
}

D:\devel\gcc\gcc>


which is stripped down from:

mode_multiple = GET_MODE_SIZE (xmode) / GET_MODE_SIZE (ymode);
if (mode_multiple == 0)

in function:

unsigned int
subreg_regno_offset (xregno, xmode, offset, ymode)

of rtlanal.c


And yes, I know gcc 3.2.3 is long out of support, but I'm
interested in the i370 target which no longer exists.

Any suggestions on where to stick some printfs so that
I can start looking for a divergence, given that the
generated code is meant to be identical, with the possible
exception of register selection which is dependent on
floating point calculations, and the floating point is known
to differ on my different environments? (because the PC
uses IEEE, Hercules has proper IBM hex float, and mfemul
has bastardized IBM hex float which you can see below
if interested - but unrelated to my question).

https://sourceforge.net/p/pdos/gitcode/ci/master/tree/util/mfemul.c

Thanks. Paul.