https://gcc.gnu.org/bugzilla/show_bug.cgi?id=91109
--- Comment #6 from Bernd Edlinger <bernd.edlinger at hotmail dot de> ---
with this patch the relevant part if the reload dump file looks different:
(insn 3414 6591 6682 129 (set (mem/c:SI (reg/f:SI 5 r5 [5715]) [1
s.5566D.5531+0 S4 A32])
(reg:SI 6 r6 [orig:828 _821 ] [828])) "20040709-1.c":13:5 654
{*arm_movsi_vfp}
(nil))
[...]
(insn 6826 3453 6816 129 (parallel [
(set (reg:SI 3 r3 [4187])
(truncate:SI (lshiftrt:DI (mult:DI (zero_extend:DI (reg:SI 11
fp [4186]))
(zero_extend:DI (reg:SI 7 r7 [4188])))
(const_int 32 [0x20]))))
(clobber (reg:SI 1 r1 [5970]))
]) "20040709-1.c":108:291 70 {*umulsi3_highpart_v6}
(nil))
[...]
(insn 3509 3530 3531 132 (set (mem/c:SI (reg/f:SI 5 r5 [5715]) [1
s.5566D.5531+0 S4 A32])
(reg:SI 14 lr [orig:692 D.6083 ] [692])) "20040709-1.c":13:5 654
{*arm_movsi_vfp}
(nil))
This time insn 6826 is able to choose a different register than r5,
and most importantly the live-range info is correct, since the
old register r5970 is renamed to r6374 temporarily:
Creating newreg=6374 from oldreg=5970, assigning class GENERAL_REGS to
scratch pseudo copy r6374
6816: r6364:SI=r4187:SI
REG_DEAD r4187:SI
Inserting rematerialization insn before:
6826: {r6364:SI=trunc(zero_extend(r4186:SI)*zero_extend(r4188:SI)
0>>0x20);clobber r6374:SI;}
REG_UNUSED r6374:SI
which is visible in the live ranges (which was not there before):
r6007: [59..59]
r6374: [1572..1572]
Compressing live ranges: from 3802 to 75 - 1%
Ranges after the compression:
[...]
r6007: [1..1]
r6374: [40..40]
However since the re-materialized instruction is able to use r1
there is no conflict any more. So I believe the patch is a
straight improvement over the previous state of affairs.
So, as it looks like, this is a potentially catastrophic bug, and not
related to -flto at all or any specific target architecture.
From my testing it is likely that was already there in gcc-9.0.1.
