[Bug target/50751] SH Target: Displacement addressing does not work for QImode and HImode

["oleg dot endo at t-online dot de" <gcc-bugzilla at gcc dot gnu dot org>]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=50751

--- Comment #4 from Oleg Endo <oleg.endo@t-online.de> 2011-10-23 21:56:56 UTC ---
Created attachment 25582
  --> http://gcc.gnu.org/bugzilla/attachment.cgi?id=25582
Experimental patch for mov.b with displacement addressing

> (In reply to comment #2)
> 
> Welcome to the spill-failure-for-class-'R0_REGS' club :-)

The attached patch is an experimental example only.  Please ignore wrong
formatting, comments, and the fact that it cripples some of the SH2A support :}

It adds support for 'mov.b @(disp, Rm), R0' and 'mov.b R0, @(disp, Rn)'
instructions.  I haven't tested it fully, only building the CSiBE
set for -m4-single -ml. 

Although it results in suboptimal code like ...

    ...
    mov.l    @(4,r15),r2
    mov    r1,r0
    mov.b    r0,@(5,r2)
    rts
    mov    r1,r0        ! redundant

.. or like ...

    mov.b    @(1,r4),r0
    mov    r0,r4
    mov.b    @(2,r5),r0
    add    r0,r4        ! better: add r4, r0
    mov    r4,r0        ! not needed if add operands are swapped
    mov.b    r0,@(5,r6)
    rts
    mov    r4,r0        ! redundant

.. it already shows some code size improvements:

avg: -563.222222 / -0.942376 %
max: compiler        22804 -> 22928         +124 / +0.543764 %
min: OpenTCP-1.0.4   27069 -> 25989        -1080 / -3.989804 %

top 5 files
mpeg2dec-0.3.1  libmpeg2/motion_comp 
  6044 -> 4796         -1248 / -20.648577 %

libpng-1.2.5  pngrtran
  19668 -> 18904         -764 / -3.884482 %

linux-2.4.23-pre3-testplatform  arch/testplatform/kernel/traps
  6192 -> 5532          -660 / -10.658915 %

lwip-0.5.3.preproc  src/core/tcp_input
  5424 -> 5040          -384 / -7.079646 %

libmspack  test/cabextract_md5
  21780 -> 21424         -356 / -1.634527 %


The R0 clobber in the movqi expander and the explicit usage of R0 in the 
splits effectively disable some optimizations, but this is the only 
thing I could get to work so far.
I've left the straight forward but non-working patterns as comments in the
patch as a reference.  Basically, without the R0 clobber in the movqi expander
it eventually ends up like that...

error: insn does not satisfy its constraints:
(insn 737 40 42 4 (set (reg:QI 10 r10)
        (mem/c:QI (plus:SI (reg:SI 1 r1 [386])
                (const_int 1 [0x1])) [0 *D.4946_20+0 S1 A8]))
{*movqi_m_reg_disp_load}
     (nil))

internal compiler error: in reload_cse_simplify_operands, at postreload.c:403

I'm puzzled why the register allocator ignores the constraint "z" when it
starts to run out of registers.  In the error case above it tries to produce
something like 'mov.b @(1,r1),r10' which of course is impossible.
Any hints are highly appreciated.