[maarten-boot]

I compiled random C programs (integer arithmetic of all widths,
conversions, comparisons, guarded division, shifts, loops, switches,
arrays, structs) with gcc and with the ACK at -O0 to -O4 and compared
the output. These are the differences that are not specific to my
target; each commit has a test in tests/plat/core that fails without
it.

• cemcom.ansi: unary - of an unsigned constant was not cut to the
  type's size (~ was), so -(unsigned long)1371337135 / 193754694
  folded to 717414569 instead of 15. Fails on all four tested
  platforms.
• em_opt:
  • repl_mul (-ml) kept the multiplier in a 64-bit unsigned long: x * 0xFFFFFFFFUL became a shift by 64 and gave 0 on
    linux68k.
  • loc loc adu/sbu/mlu/sli/slu (and the ldc forms) folded to
    constants that do not fit the word, and loc loc loc cii/ciu
    folded without sign-extending the constant ((int8_t)16384 stayed
    16384). Both show at -O3, when ego propagates constants: the tests
    fail on linux68k, linuxppc and pc86.
• m68020 ncg table (linux68k):
  • the memory1 coercion (clr, move.b) kept the condition codes
    of the byte as those of the register: (uint8_t)l / 16 with
    l = -300 gave 14.
  • x & (a op b) was done as "clear bit 0 of x if false", only right
    for x of 0 or 1: (x >> s) & (a < b) with x = -8017 gave -8017.

Tested: make with PLATS = all, all tests pass on cpm, linux68k,
linuxppc and pc86; without the fixes the new tests fail as described.
Independent of #381 and #382.

🤖 Generated with Claude Code

[davidgiven]

Merging; thanks.