[Bug target/53949] [SH] Add support for mac.w / mac.l instructions

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53949

--- Comment #13 from Oleg Endo <olegendo at gcc dot gnu.org> ---
A more interesting real-world example from libjpeg would be function
jpeg_idct_ifast (jidctint.c).

If we take the code as-is, there are few mac opportunities due to sharing of
the terms.  The expressions could be un-CSE'd which would result in longer mac
chains, but the overall result gets worse because the data layout is not in a
mac friendly way.

The first loop in jpeg_idct_ifast can be split into 8 independent loops for the
output value wsptr[8*n+i].
For n = 1,2,3,4,5,6 the loops look a bit complex, but for n = 0 and n = 7 we
get similar looking loops like:

for (int i = 0; i < 8; ++i)
{
  wsptr[8*7+i] = inptr[8*0 + i] * quantptr[8*0 + i]
                 - inptr[8*1 + i] * quantptr[8*1 + i]
                 + inptr[8*2 + i] * quantptr[8*2 + i]
                 - inptr[8*3 + i] * quantptr[8*3 + i]
                 + inptr[8*4 + i] * quantptr[8*4 + i]
                 - inptr[8*5 + i] * quantptr[8*5 + i]
                 + inptr[8*6 + i] * quantptr[8*6 + i]
                 - inptr[8*7 + i] * quantptr[8*7 + i];
}

Still, due to the subtractions and memory access pattern, plain mac insns can't
be used.

The subtractions can be converted into additions by negating the operands. 
Since mac wants both operands in memory, those can be placed on the stack. 
Also, in this case the address registers can be pre-computed outside the loop,
since there are enough registers.
A possible outcome would be something like this:

                                    // r4 = inptr[8*0+i]
                                    // r5 = quantptr[8*0+i]
                                    // r6 = wsptr[8*0+i]

  mov    r4,r3;    add    #32,r3    // r3 = inptr[8*1+i]
  mov    r3,r7;    add    #32,r7    // r7 = inptr[8*2+i]
  mov    r7,r8;    add    #32,r8    // r8 = inptr[8*3+i]
  mov    r8,r9;    add    #32,r9    // r9 = inptr[8*4+i]
  mov    r9,r10;   add    #32,r10   // r10 = inptr[8*5+i]
  mov    r10,r11;  add    #32,r11   // r11 = inptr[8*6+i]
  mov    r11,r12;  add    #32,r12   // r12 = inptr[8*7+i]

  mov    #8,r14
  add    #126,r6;  add    #102,r6   // r6 = wpstr + 8*7*4 + 4
  mov    r4,r0;    sub    r5,r0     // r0 = quantptr - intptr

.Loop:
  mov.l  @(r0,r12),r1   // quantptr[8*7+i]
  mov.l  @(r0,r11),r2    // quantptr[8*6+i]
  mov.l  @(r0,r10),r13    // quantptr[8*5+i]
  neg    r1,r1
  mov.l  r1,@-r15
  mov.l  r2,@-r15
  neg    r13,r13
  mov.l  @(r0,r8),r1    // quantptr[8*3+i]
  mov.l  @(r0,r9),r2    // quantptr[8*4+i]
  mov.l  r13,@-r15
  neg    r1,r1
  mov.l  r2,@-r15
  mov.l  @(r0,r7),r2    // quantptr[8*2+i]
  mov.l  @(r0,r3),r13   // quantptr[8*1+i]
  mov.l  r1,@-r15
  mov.l  r2,@-r15
  neg    r13,r13
  mov.l  r13,@-r15

  clrmac
  mac.l    @r4+,@r5+
  mac.l    @r3+,@r15+
  mac.l    @r7+,@r15+
  mac.l    @r8+,@r15+
  mac.l    @r9+,@r15+
  mac.l    @r10+,@r15+
  mac.l    @r11+,@r15+
  mac.l    @r12+,@r15+
  dt    r14
  sts    macl,@-r6
  bf/s    .Loop
  add    #8,r6

which is 31 insns per loop and (almost) no pipeline stalls, vs. 53 insns per
loop + stalls on mul-sts sequences when the mac insn is not used.
The above loop can be optimized even further with partial unrolling to avoid
the latency of the last mac and sts.
Of course it'd be even better, if the application's data was in a mac friendly
layout.