[Bug target/111829] Redudant register moves inside the loop

[crazylht at gmail dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=111829

--- Comment #3 from Hongtao.liu <crazylht at gmail dot com> ---
(In reply to Richard Biener from comment #2)
> You sink the conversion, so it would be PRE on the reverse graph.  The
> transform doesn't really fit a particular pass I think.
> 
> Why does the problem persist in RTL?
Normally, combine will eliminate the redudant move by combine subreg to the
pattern like.

1004(insn 19 17 21 3 (set (subreg:V4SI (reg/v:V2DI 103 [ vsum ]) 0)
1005        (unspec:V4SI [
1006                (subreg:V4SI (reg/v:V2DI 103 [ vsum ]) 0)
1007                (reg:V4SI 123 [ MEM[(__m128i * {ref-all})_52] ])
1008                (reg:V4SI 124)
1009            ] UNSPEC_VPDPBUSD)) "test.c":9:16 discrim 1 9182
{vpdpbusd_v4si}

but for this case, before combine, cse1/fwprop propagate the subreg(insn 21)
from inner loop to outside(insn 28), since there's use for (reg:V4SI 121),
combine failed to eliminate the redudnat mov of subreg.

------loop_begin----------
...
(insn 19 18 20 3 (set (reg:V4SI 121)
393        (unspec:V4SI [
394                (reg:V4SI 122 [ vsum ])
395                (reg:V4SI 123 [ MEM[(__m128i * {ref-all})_52] ])
396                (reg:V4SI 124)
397            ] UNSPEC_VPDPBUSD)) "test.c":9:16 discrim 1 9182 {vpdpbusd_v4si}
398     (expr_list:REG_DEAD (reg:V4SI 125)
399        (expr_list:REG_DEAD (reg:V4SI 123 [ MEM[(__m128i * {ref-all})_52] ])
400            (expr_list:REG_DEAD (reg:V4SI 122 [ vsum ])
401                (nil)))))
402(insn 20 19 21 3 (set (reg:V4SI 102 [ _11 ])
403        (reg:V4SI 121)) "test.c":9:16 discrim 1 1906 {movv4si_internal}
404     (expr_list:REG_DEAD (reg:V4SI 121)
405        (nil)))
406(insn 21 20 22 3 (set (reg/v:V2DI 103 [ vsum ])
407        (subreg:V2DI (reg:V4SI 121) 0)) "test.c":9:16 discrim 2 1909
{movv2di_internal}
408     (nil))
...
---------loop_end---------

453(note 27 26 28 4 [bb 4] NOTE_INSN_BASIC_BLOCK)
454(insn 28 27 29 4 (set (mem:V2DI (reg/v/f:DI 119 [ pc ]) [0 *pc_22(D)+0 S16
A128])
455        (subreg:V2DI (reg:V4SI 121) 0)) "test.c":11:9 1909
{movv2di_internal}
456     (expr_list:REG_DEAD (reg/v/f:DI 119 [ pc ]) --- propogate from insn 21
457        (expr_list:REG_DEAD (reg/v:V2DI 103 [ vsum ])
458            (nil))))