[Bug target/29560] [avr] Poor optimization for byte shifts

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

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=29560

--- Comment #7 from Georg-Johann Lay <gjl at gcc dot gnu.org> 2011-07-23 14:50:58 UTC ---
Created attachment 24816
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=24816
Test case for 16-bit shifts that use low part only

Reopened this optimization issue.

For the attached test case, there are sometimes REG_UNUSED notes for the high
part like with avr-gcc-4.6.1 -Os -dP

shift1:
 ; (insn 8 4 17 (set (reg:HI 24 r24 [50])
 ;         (ashift:HI (reg:HI 24 r24 [ x ])
 ;             (reg/v:QI 22 r22 [orig:47 s ] [47]))) foo.c:3 68 {ashlhi3}
 ;      (expr_list:REG_DEAD (reg/v:QI 22 r22 [orig:47 s ] [47])
 ;         (expr_list:REG_UNUSED (reg:QI 25 r25)
 ;             (nil))))
    rjmp 2f     ;  8    ashlhi3/1    [length = 6]
1:    lsl r24
    rol r25
2:    dec r22
    brpl 1b

So that there is a way to map ashlhi3 to ashlqi3, i.e. 16-bit shift to a 8-bit
shift.

Unfortunately, these notes are not always present like in shift2:

shift2:
 ; (insn 15 4 8 (set (reg:HI 18 r18)
 ;         (reg:HI 24 r24 [ x ])) foo.c:10 10 {*movhi}
 ;      (nil))
    movw r18,r24     ;  15    *movhi/1    [length = 1]
 ; (insn 8 15 9 (set (reg:HI 18 r18)
 ;         (ashift:HI (reg:HI 18 r18)
 ;             (reg/v:QI 22 r22 [orig:46 s ] [46]))) foo.c:10 68 {ashlhi3}
 ;      (expr_list:REG_DEAD (reg/v:QI 22 r22 [orig:46 s ] [46])
 ;         (nil)))
    rjmp 2f     ;  8    ashlhi3/1    [length = 6]
1:    lsl r18
    rol r19
2:    dec r22
    brpl 1b

The notes are not present in pass .split2 so a split won't help. The notes
appear to be available in .peephole2 so that could be a fix. Moreover, the
notes won't be back-propagated so that an optimization will cover at most one
insn: the shift insn.