unneeded spills of SF values on xtensa with FPU

[Max Filippov]

Hi Suwa-san,

I've noticed that in xtensa configurations with hardware FPU
function arguments of type float are spilled on the stack although
there's no need for that. E.g. the following function:

int f(float a, float b)
{
    return a < b;
}

translates to the following with -O2:

f:
       entry   sp, 48
       wfr     f0, a2
       wfr     f1, a3
       s32i.n  a2, sp, 0
       olt.s   b0, f0, f1
       movi.n  a8, 0
       movi.n  a2, 1
       s32i.n  a3, sp, 4
       movf    a2, a8, b0
       retw.n

The relevant RTL looks like this at the end of IRA:

(insn 18 4 19 2 (set (reg:SF 51)
       (reg:SF 2 a2 [ a ])) "test2.c":2:1 61 {movsf_internal}
    (expr_list:REG_DEAD (reg:SF 2 a2 [ a ])
       (nil)))
(insn 19 18 7 2 (set (reg:SF 52)
       (reg:SF 3 a3 [ b ])) "test2.c":2:1 61 {movsf_internal}
    (expr_list:REG_DEAD (reg:SF 3 a3 [ b ])
       (nil)))
(insn 7 19 21 2 (set (reg:CC 18 b0)
       (lt:CC (reg:SF 51)
           (reg:SF 52))) "test2.c":3:11 100 {slt_sf}
    (expr_list:REG_DEAD (reg:SF 52)
       (expr_list:REG_DEAD (reg:SF 51)
           (nil))))

and it is transformed to the following by the end of LRA:

(insn 18 4 19 2 (set (mem/c:SF (reg/f:SI 1 sp) [1 %sfp+0 S4 A32])
       (reg:SF 2 a2 [ a ])) "test2.c":2:1 61 {movsf_internal}
    (nil))
(insn 19 18 24 2 (set (mem/c:SF (plus:SI (reg/f:SI 1 sp)
               (const_int 4 [0x4])) [1 %sfp+4 S4 A32])
       (reg:SF 3 a3 [ b ])) "test2.c":2:1 61 {movsf_internal}
    (nil))
(insn 24 19 25 2 (set (reg:SF 19 f0 [51])
       (mem/c:SF (reg/f:SI 1 sp) [1 %sfp+0 S4 A32])) "test2.c":3:11 61
{movsf_internal}
    (nil))
(insn 25 24 7 2 (set (reg:SF 20 f1 [52])
       (mem/c:SF (plus:SI (reg/f:SI 1 sp)
               (const_int 4 [0x4])) [1 %sfp+4 S4 A32])) "test2.c":3:11
61 {movsf_internal}
    (nil))
(insn 7 25 21 2 (set (reg:CC 18 b0)
       (lt:CC (reg:SF 19 f0 [51])
           (reg:SF 20 f1 [52]))) "test2.c":3:11 100 {slt_sf}
    (nil))

LRA stops checking alternatives for insns 18 and 19 at s32i.n,
but even if I move wfr at the head of the movsf_internal list it
still loses to s32i.n.
Postreload pass replaces the lsi instructions 24 and 25 with
wfr from a2 and a3, but doesn't remove the spills.
I wonder what can be done with that?

--
Thanks.
-- Max