[Bug target/64345] [SH] Improve single bit extraction

["olegendo at gcc dot gnu.org" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=64345

--- Comment #2 from Oleg Endo <olegendo at gcc dot gnu.org> ---
A recent change in the middle end has triggered this:

FAIL: gcc.target/sh/pr54236-1.c scan-assembler-times negc 2

The test

int
test_07 (int *vec)
{
  /* Must not see a 'sett' or 'addc' here.
     This is a case where combine tries to produce
     'a + (0 - b) + 1' out of 'a - b + 1'.
     On non-SH2A there is a 'tst + negc', on SH2A a 'bld + movt'.  */
  int z = vec[0];
  int vi = vec[1];
  int zi = vec[2];

  if (zi != 0 && z < -1)
    vi -= (((vi >> 7) & 0x01) << 1) - 1;

  return vi;
}

fails to produce the expected tst + negc sequence.
A reduced case is:

int test (int vi)
{
  return vi - (((vi >> 6) & 0x01) << 1);
}

-m2a -O2 (gcc 5):
        bld     #6,r4
        movt    r0
        rts     
        add     r0,r0

trunk:
        mov     #-5,r1
        mov     r4,r0
        shld    r1,r0
        rts     
        and     #2,r0

-m4 -O2 (gcc 5):
        mov     r4,r0
        tst     #64,r0
        mov     #-1,r0
        negc    r0,r0
        rts     
        add     r0,r0

trunk:
        mov     #-5,r1
        mov     r4,r0
        shld    r1,r0
        rts     
        and     #2,r0


For the more complex expression

int test (int vi)
{
  return vi - (((vi >> 6) & 0x01) << 1) - 1;
}

-m4 -O2 (gcc 5):
        mov     r4,r0
        mov     #-1,r1
        tst     #64,r0
        negc    r1,r1
        add     r1,r1
        rts     
        subc    r1,r0

trunk:
        mov     #-5,r1
        mov     r4,r0
        shld    r1,r0
        sett
        and     #2,r0
        subc    r0,r4
        rts     
        mov     r4,r0

It seems the zero_extract patterns are not used for this anymore.  Instead
combine is looking for something like

(set (reg:SI 168 [ D.1649 ])
    (and:SI (lshiftrt:SI (reg:SI 4 r4 [ vi ])
            (const_int 5 [0x5]))
        (const_int 2 [0x2])))

This pattern should be added as a zero_extract variant.