[Bug target/53987] [SH] Unnecessary zero-extensions

olegendo at gcc dot gnu.org gcc-bugzilla@gcc.gnu.org
Thu Jan 15 01:37:00 GMT 2015


https://gcc.gnu.org/bugzilla/show_bug.cgi?id=53987

--- Comment #9 from Oleg Endo <olegendo at gcc dot gnu.org> ---
Another related issue is this example function:

unsigned char h (unsigned char a, int b)
{
  return (unsigned char)(a + b);
}

compiling with -O2 -mrenesas (which allows undefined high bits in the return
value):
        extu.b  r4,r4    << not needed
        mov     r5,r0
        rts
        add     r4,r0

It seems that the fwprop1 pass causes this.
The initial RTL for this looks like this:

(insn 2 5 3 2 (set (reg/v:SI 165 [ a ])
        (zero_extend:SI (reg:QI 4 r4 [ a ]))) sh_tmp.cpp:1550 -1
     (nil))
(insn 3 2 4 2 (set (reg/v:SI 166 [ b ])
        (reg:SI 5 r5 [ b ])) sh_tmp.cpp:1550 -1
     (nil))
(note 4 3 7 2 NOTE_INSN_FUNCTION_BEG)
(insn 7 4 8 2 (set (reg:QI 168)
        (subreg:QI (reg/v:SI 166 [ b ]) 0)) sh_tmp.cpp:1551 -1
     (nil))
(insn 8 7 9 2 (set (reg:SI 169)
        (plus:SI (reg/v:SI 165 [ a ])
            (subreg:SI (reg:QI 168) 0))) sh_tmp.cpp:1551 -1
     (nil))
(insn 9 8 13 2 (set (reg:QI 164 [ <retval> ])
        (subreg:QI (reg:SI 169) 0)) sh_tmp.cpp:1551 -1
     (nil))
(insn 13 9 14 2 (set (reg/i:QI 0 r0)
        (reg:QI 164 [ <retval> ])) sh_tmp.cpp:1552 -1
     (nil))
(insn 14 13 0 2 (use (reg/i:QI 0 r0)) sh_tmp.cpp:1552 -1
     (nil))


Notice that the plus op is expanded as a subreg:SI (reg:QI) and the result is
then taken as a subreg:QI.  This should be OK with the addsi3 pattern. 
However, fwprop1 then propagates the hardregs and eliminates the subregs:

(note 5 0 2 2 [bb 2] NOTE_INSN_BASIC_BLOCK)
(insn 2 5 3 2 (set (reg/v:SI 165 [ a ])
        (zero_extend:SI (reg:QI 4 r4 [ a ]))) sh_tmp.cpp:1550 218
{*zero_extendqisi2_compact}
     (expr_list:REG_DEAD (reg:QI 4 r4 [ a ])
        (nil)))
(insn 3 2 4 2 (set (reg/v:SI 166 [ b ])
        (reg:SI 5 r5 [ b ])) sh_tmp.cpp:1550 252 {movsi_ie}
     (expr_list:REG_DEAD (reg:SI 5 r5 [ b ])
        (nil)))
(note 4 3 8 2 NOTE_INSN_FUNCTION_BEG)
(insn 8 4 13 2 (set (reg:SI 169)
        (plus:SI (reg/v:SI 165 [ a ])
            (reg/v:SI 166 [ b ]))) sh_tmp.cpp:1551 63 {*addsi3_compact}
     (expr_list:REG_DEAD (reg:QI 168 [ b ])
        (expr_list:REG_DEAD (reg/v:SI 165 [ a ])
            (nil))))
(insn 13 8 14 2 (set (reg/i:QI 0 r0)
        (subreg:QI (reg:SI 169) 0)) sh_tmp.cpp:1552 270 {*movqi}
     (expr_list:REG_DEAD (reg:SI 169)
        (nil)))
(insn 14 13 0 2 (use (reg/i:QI 0 r0)) sh_tmp.cpp:1552 -1
     (nil))

After that, everything is pretty much set and the zero extension will not be
eliminated anymore.



More information about the Gcc-bugs mailing list