[Bug rtl-optimization/100342] [10/11 Regression] wrong code with -O2 -fno-dse -fno-forward-propagate -mno-sse2

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

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=100342

Jakub Jelinek <jakub at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |liuhongt at gcc dot gnu.org

--- Comment #11 from Jakub Jelinek <jakub at gcc dot gnu.org> ---
(insn 2741 1965 368 2 (set (reg:DI 42 r14 [orig:2067 u128_0 ] [2067])
        (mem/c:DI (plus:DI (reg/f:DI 7 sp)
                (const_int 56 [0x38])) [9 %sfp+-1032 S8 A64]))
"pr100342.c":65:24 66 {*movdi_internal}
     (nil))
...
(insn 613 612 1970 2 (set (reg:QI 4 si [orig:125 _98 ] [125])
        (reg:QI 42 r14 [orig:2067 u128_0 ] [2067])) "pr100342.c":68:15 69
{*movqi_internal}
     (nil))
...
(insn 2737 1971 615 2 (set (reg:DI 6 bp [orig:2067 u128_0 ] [2067])
        (reg:DI 42 r14 [orig:2067 u128_0 ] [2067])) "pr100342.c":68:12 66
{*movdi_internal}
     (nil))
...
(insn 2829 2749 2855 2 (set (reg:DI 42 r14 [2222])
        (const_int -71776119061217281 [0xff00ffffffffffff])) "pr100342.c":68:12
66 {*movdi_internal}
     (nil))
...
(insn 2275 658 659 2 (set (reg:DI 1 dx [2225])
        (reg:DI 4 si [orig:125 _98 ] [125])) "pr100342.c":68:12 66
{*movdi_internal}
     (nil))
...
(insn 2910 2860 677 2 (set (reg:DI 4 si [2227])
        (const_int -1095216660481 [0xffffff00ffffffff])) "pr100342.c":68:12 66
{*movdi_internal}
     (nil))
...
(insn 2287 2286 710 2 (set (reg:DI 1 dx [orig:131 _104 ] [131])
        (reg:DI 6 bp [orig:2067 u128_0 ] [2067])) "pr100342.c":72:5 66
{*movdi_internal}
     (nil))
(insn 710 2287 2172 2 (set (reg:DI 1 dx [orig:131 _104 ] [131])
        (bswap:DI (reg:DI 1 dx [orig:131 _104 ] [131]))) "pr100342.c":72:5 872
{*bswapdi2}
     (nil))
In copyprop_hardreg_forward_1 at the start of that insn 2737 we have:
(gdb) p vd->e[1]
$32 = {mode = E_DImode, oldest_regno = 1, next_regno = 4294967295,
debug_insn_changes = 0x0}
(gdb) p vd->e[4]
$33 = {mode = E_QImode, oldest_regno = 42, next_regno = 4294967295,
debug_insn_changes = 0x0}
(gdb) p vd->e[6]
$34 = {mode = E_SImode, oldest_regno = 6, next_regno = 4294967295,
debug_insn_changes = 0x0}
(gdb) p vd->e[42]
$35 = {mode = E_DImode, oldest_regno = 42, next_regno = 4, debug_insn_changes =
0x0}
which is I think ok, it says %sil is QImode copy of %r14, all other regs we
care about here contain their values.
At the end of that we have:
(gdb) p vd->e[1]
$36 = {mode = E_DImode, oldest_regno = 1, next_regno = 4294967295,
debug_insn_changes = 0x0}
(gdb) p vd->e[4]
$37 = {mode = E_QImode, oldest_regno = 42, next_regno = 6, debug_insn_changes =
0x0}
(gdb) p vd->e[6]
$38 = {mode = E_DImode, oldest_regno = 42, next_regno = 4294967295,
debug_insn_changes = 0x0}
(gdb) p vd->e[42]
$39 = {mode = E_DImode, oldest_regno = 42, next_regno = 4, debug_insn_changes =
0x0}
as well as before the insn 2829.  That changes it to:
(gdb) p vd->e[1]
$51 = {mode = E_DImode, oldest_regno = 1, next_regno = 4294967295,
debug_insn_changes = 0x0}
(gdb) p vd->e[4]
$52 = {mode = E_QImode, oldest_regno = 4, next_regno = 6, debug_insn_changes =
0x0}
(gdb) p vd->e[6]
$53 = {mode = E_DImode, oldest_regno = 4, next_regno = 4294967295,
debug_insn_changes = 0x0}
(gdb) p vd->e[42]
$54 = {mode = E_DImode, oldest_regno = 42, next_regno = 4294967295,
debug_insn_changes = 0x0}
Same thing at the start of insn 2275.  But after that we have:
(gdb) p vd->e[1]
$61 = {mode = E_DImode, oldest_regno = 4, next_regno = 4294967295,
debug_insn_changes = 0x0}
(gdb) p vd->e[4]
$62 = {mode = E_QImode, oldest_regno = 4, next_regno = 6, debug_insn_changes =
0x0}
(gdb) p vd->e[6]
$63 = {mode = E_DImode, oldest_regno = 4, next_regno = 1, debug_insn_changes =
0x0}
(gdb) p vd->e[42]
$64 = {mode = E_DImode, oldest_regno = 42, next_regno = 4294967295,
debug_insn_changes = 0x0}
I think this step is what looks wrong, it is true that the movq %rsi, %rdx
copied all 64-bits of %rsi, but vd->e[4].mode was just QImode, so I think
vd->e[1].mode
should be also QImode.
That remains until before insn 2910, which changes that to:
(gdb) p vd->e[1]
$84 = {mode = E_DImode, oldest_regno = 6, next_regno = 4294967295,
debug_insn_changes = 0x0}
(gdb) p vd->e[4]
$85 = {mode = E_DImode, oldest_regno = 4, next_regno = 4294967295,
debug_insn_changes = 0x0}
(gdb) p vd->e[6]
$86 = {mode = E_DImode, oldest_regno = 6, next_regno = 1, debug_insn_changes =
0x0}
(gdb) p vd->e[42]
$87 = {mode = E_DImode, oldest_regno = 42, next_regno = 4294967295,
debug_insn_changes = 0x0}
and at that point, the fact that %rdx doesn't contain the same value as %rbp -
only
%dl is guaranteed to be the same as %bpl - is lost.
Then when we look up the oldest value of %rdx on insn 2287 we are told it is
%rbp.