Hi DJ,
I managed to reproduce the issue, it is in function test_7, sorry I'm not
sure how I missed it I always compare the logs carefully with and without
the patch.
After investigating this I found the problem is caused by a removal of 1
instruction due to wrong UNUSED note being
added(rl78_calculate_death_notes), more exactly in rl78_note_reg_set :
r = REGNO (d);
if (dead [r])
add_reg_note (insn, REG_UNUSED, gen_rtx_REG (GET_MODE (d), r));
For example if the reg is HImode and the high part is not dead (dead[r + 1])
the register should NOT be marked as UNUSED but it is.
After I corrected the problem the code was generated correctly; I will
prepare another patch with this fix.
Below you can see the relevant parts from the rtl dump(.314r.devirt) which
show this:
(insn 41 91 92 2 (set (reg:HI 0 x)
(mem/c:HI (reg:HI 6 l) [0 S2 A16])) "../test_movdi.c":20 43
{*movhi_real}
(nil))
(insn 92 41 42 2 (set (reg:HI 8 r8 [orig:44 a.0_3 ] [44])
(reg:HI 0 x)) "../test_movdi.c":20 -1
(nil))
...
(insn 17 105 106 2 (set (reg:QI 1 a)
(reg:QI 9 r9 [ a.0_3+1 ])) "../test_movdi.c":20 42 {*movqi_real}
(nil))
...
(insn 19 107 108 2 (set (reg:QI 1 a)
(reg:QI 9 r9 [ a.0_3+1 ])) "../test_movdi.c":20 42 {*movqi_real}
(nil))
...
(insn 22 110 23 2 (set (reg/f:HI 8 r8 [53])
(symbol_ref:HI ("c") <var_decl 0xb713c0fc c>)) "../test_movdi.c":20
43 {*movhi_real}
(expr_list:REG_EQUIV (symbol_ref:HI ("c") <var_decl 0xb713c0fc c>)
(nil)))
When the death notes are calculated we have the following
(rl78_calculate_death_notes):
Because this parse the functions backwards insn 22 is processed first.
note set reg 8 size 2
(insn 22 110 23 2 (set (reg/f:HI 8 r8 [53])
(symbol_ref:HI ("c") <var_decl 0xb713c0fc c>)) "../test_movdi.c":20
43 {*movhi_real}
(expr_list:REG_EQUIV (symbol_ref:HI ("c") <var_decl 0xb713c0fc c>)
(nil)))
--------------------------------------------------
Dead: x a c b r8 r9 r10 r11 r12 r13 r16 r17 <-----------As you can see R8
and R9 are marked as dead which is correct
....
note set reg 1 size 1
note use reg 9 size 1 on insn 19
(reg:QI 9 r9 [ a.0_3+1 ])
(insn 19 107 108 2 (set (reg:QI 1 a)
(reg:QI 9 r9 [ a.0_3+1 ])) "../test_movdi.c":20 42 {*movqi_real}
(expr_list:REG_DEAD (reg:QI 9 r9)
(nil)))
--------------------------------------------------
Dead: x a c b r8 r10 r13 r16 r17 <-----------As you can see R9 is not dead
anymore
....
Now we get to insn 92 and we can see that r8 is wrongfully marked as unused
because the code check only R8 (and it should check R9 as well).
As a consequence insn 92 get removed later on.
--------------------------------------------------
Dead: x a c b r8 r10 r11 r12 r13 r14 r15 r16 r17
(insn 92 41 42 2 (set (reg:HI 8 r8 [orig:44 a.0_3 ] [44])
(reg:HI 0 x)) "../test_movdi.c":20 43 {*movhi_real}
(nil))
note set reg 8 size 2
note use reg 0 size 2 on insn 92
(reg:HI 0 x)
(insn 92 41 42 2 (set (reg:HI 8 r8 [orig:44 a.0_3 ] [44])
(reg:HI 0 x)) "../test_movdi.c":20 43 {*movhi_real}
(expr_list:REG_DEAD (reg:HI 0 x)
(expr_list:REG_UNUSED (reg:HI 8
r8)<------------------------------------here is the problem
(nil))))
Also later on (in 290r.peephole2) we see that instruction 41 is eliminated
as well:
DCE: Deleting insn 41
deleting insn with uid = 41.
Best Regards,
Sebastian
> -----Original Message-----
> From: DJ Delorie [mailto:d...@redhat.com]
> Sent: 09 January 2018 22:55
> To: Sebastian Perta <sebastian.pe...@renesas.com>
> Cc: gcc-patches@gcc.gnu.org
> Subject: Re: [PATCH] RL78 movdi improvement
>
>
> I saw one regression with this patch:
>
> PASS-FAIL: gcc.dg/torture/vshuf-v8qi.c -O2 execution test
>
> Could you confirm this? Note: I'm testing with your other (pre-2018 at
> least) patches applied at the same time (smax etc, anddi, bswaphi) so it
> might be an interaction, but the code looks like a movdi bug. The other
> patches look OK otherwise.
