[Bug rtl-optimization/70961] Regrename ignores preferred_rename_class

["wdijkstr at arm dot com" <gcc-bugzilla at gcc dot gnu dot org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=70961

--- Comment #5 from Wilco <wdijkstr at arm dot com> ---
As for a simple example, Proc_4 in Dhrystone is a good one. With -O2 and
-fno-rename-registers I get the following on Thumb-2:

000000c8 <Proc_4>:
  c8:   b430            push    {r4, r5}
  ca:   f240 0300       movw    r3, #0
  ce:   f240 0400       movw    r4, #0
  d2:   f2c0 0300       movt    r3, #0
  d6:   f2c0 0400       movt    r4, #0
  da:   f240 0100       movw    r1, #0
  de:   681a            ldr     r2, [r3, #0]
  e0:   f2c0 0100       movt    r1, #0
  e4:   7824            ldrb    r4, [r4, #0]
  e6:   2542            movs    r5, #66 ; 0x42
  e8:   700d            strb    r5, [r1, #0]
  ea:   2c41            cmp     r4, #65 ; 0x41
  ec:   bf08            it      eq
  ee:   f042 0201       orreq.w r2, r2, #1
  f2:   601a            str     r2, [r3, #0]
  f4:   bc30            pop     {r4, r5}
  f6:   4770            bx      lr

With -frename-registers:

000000c8 <Proc_4>:
  c8:   b430            push    {r4, r5}
  ca:   f240 0300       movw    r3, #0
  ce:   f240 0400       movw    r4, #0
  d2:   f2c0 0300       movt    r3, #0
  d6:   f2c0 0400       movt    r4, #0
  da:   f240 0100       movw    r1, #0
  de:   681a            ldr     r2, [r3, #0]
  e0:   f2c0 0100       movt    r1, #0
  e4:   f894 c000       ldrb.w  ip, [r4]
  e8:   2542            movs    r5, #66 ; 0x42
  ea:   700d            strb    r5, [r1, #0]
  ec:   f1bc 0f41       cmp.w   ip, #65 ; 0x41
  f0:   bf08            it      eq
  f2:   f042 0201       orreq.w r2, r2, #1
  f6:   601a            str     r2, [r3, #0]
  f8:   bc30            pop     {r4, r5}
  fa:   4770            bx      lr


So here it changed 2 16-bit instructions into 32-bit ones by changing R4 to IP.
However there is no benefit in doing so as it doesn't remove a move or create
additionally scheduling opportunities (the first liverange of R4 has only a
single use, so when it goes dead it is actually best to reuse R4 rather than
use a different register).