[Bug target/84154] New: PowerPC GCC 7 and 8 have regression in converting fp to short/char and returning it

[meissner at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=84154

            Bug ID: 84154
           Summary: PowerPC GCC 7 and 8 have regression in converting fp
                    to short/char and returning it
           Product: gcc
           Version: 8.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: target
          Assignee: unassigned at gcc dot gnu.org
          Reporter: meissner at gcc dot gnu.org
  Target Milestone: ---

In preparing to close target/79038, I decided to run some tests.  I discovered
that both GCC 8 and GCC 7 regress compared to GCC 6 if you have a function that
returns a short/char value and the value is converted from floating point when
you are compiling for a target with direct move (i.e. power8/power9).

In particular, if you have a function of the form:
unsigned char d2u (double x) { return x; }

it will generate the following code for power8:
        fctiwuz 0,1
        mfvsrwz 9,0
        stw 9,-16(1)
        ori 2,2,0
        lbz 3,-16(1)
        blr

I.e., it will generate a store and then a load.  GCC 6 would just do the zero
extend inline:        fctiwuz 1,1
        mfvsrd 3,1
        rlwinm 3,3,0,0xff
        blr

For -mcpu=power8, the following combinations generate store/load (double to
signed char is ok):
       double -> unsigned char
       double -> unsigned short
       double -> signed short

For -mcpu=power9, the following combinations generate store/load (double to all
char/short is ok, also _Float128 to signed char is ok):
       _Float128 -> unsigned char
       _Float128 -> unsigned short
       _Float128 -> signed short