https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120459
Jeffrey A. Law <law at gcc dot gnu.org> changed:
What |Removed |Added
----------------------------------------------------------------------------
Resolution|--- |DUPLICATE
Status|UNCONFIRMED |RESOLVED
--- Comment #5 from Jeffrey A. Law <law at gcc dot gnu.org> ---
Same core issue. We have this in IRA:
(insn 527 536 655 6 (set (reg:DF 276 [ _215 ])
(float:DF (reg:DI 1174 [ MEM[(long int *)&bi + 16B] ]))) "j.c":101:55
167 {floatdidf2}
(expr_list:REG_DEAD (reg:DI 1174 [ MEM[(long int *)&bi + 16B] ])
(nil)))
(reg 1174) does not get a hard register in the context where this use appears.
So we turn (reg 1174) into a (mem (fp) (offset))
In the .reload dump:
Changing pseudo 1174 in operand 1 of insn 527 on equiv [frame:DI+0x18]
[ ... ]
Inserting insn reload before:
1693: r2060:DI=[frame:DI+0x18]
LRA iterates to do final resolutions. Initially the offset fits, but as
further spills occur the offset no longer fits:
1693: r2060:DI=[r2652:DI]
Inserting insn reload before:
2294: r2652:DI=frame:DI+0x808
Giving us:
(insn 2294 536 1693 6 (set (reg:DI 10 a0 [2652])
(plus:DI (reg/f:DI 2 sp)
(const_int 2056 [0x808]))) "j.c":101:55 7
{*adddi3_const_sum_of_two_s12}
(nil))
(insn 1693 2294 527 6 (set (reg:DI 9 s1 [2060])
(mem:DI (reg:DI 10 a0 [2652]) [3 MEM[(long int *)&bi + 16B]+0 S8 A64]))
"j.c":101:55 277 {*movdi_64bit}
(nil))
(insn 527 1693 1694 6 (set (reg:DF 38 ft6 [orig:276 _215 ] [276])
(float:DF (reg:DI 9 s1 [2060]))) "j.c":101:55 167 {floatdidf2}
(nil))
Where insn 2294 will get split creating the scenario reported by Paul-Antoine.
The testcase in the more recently filed bug is easier to follow. So we'll use
that bug going forward.
*** This bug has been marked as a duplicate of bug 120811 ***
