Hi,
The attached patch tries to fix PR88833.
For the following test-case:

subroutine foo(x)
  real :: x(100)
  x = x + 10
end subroutine foo

Assembly with -O3 -march=armv8.2-a+sve:

foo_:
.LFB0:
        .cfi_startproc
        mov     w2, 100
        mov     w3, w2
        mov     x1, 0
        whilelo p0.s, wzr, w2
        fmov    z1.s, #1.0e+1
        .p2align 3,,7
.L2:
        ld1w    z0.s, p0/z, [x0, x1, lsl 2]
        fadd    z0.s, z0.s, z1.s
        st1w    z0.s, p0, [x0, x1, lsl 2]
        incw    x1
        whilelo p0.s, w1, w3
        bne     .L2
        ret

As we can see, it generates extra mov w3, w2. Instead it could have reused w2 in
both whilelo's.

expand produces:
insn 7: reg:SI 97 = 100
insn 8: use (reg:SI 97)
insn 22: reg:SI 105 = 100
insn 23: use (reg:SI 105)

Both reg:SI 97 and reg:SI 105 have only single definitions (and also
single use).

cse2 then replaces 100 with reg:SI 97 in insn 22, which becomes:
insn 22: reg:SI 105 = reg:SI 97.

sched1 then reorders instructions, and insn 7 and insn 22 end up falling
in same basic block Looking at reload dump:

         Choosing alt 3 in insn 7:  (0) r  (1) M {*movsi_aarch64}
          alt=0,overall=0,losers=0,rld_nregs=0
         Choosing alt 0 in insn 2:  (0) =r  (1) r {*movdi_aarch64}
          alt=0,overall=0,losers=0,rld_nregs=0
         Choosing alt 0 in insn 22:  (0) =r  (1) r {*movsi_aarch64}
            1 Non-pseudo reload: reject+=2
            1 Non input pseudo reload: reject++
            Cycle danger: overall += LRA_MAX_REJECT
          alt=0,overall=609,losers=1,rld_nregs=1

which shows, it ends up taking extra register.

The issue here is that cse2 pass is leaving opportunities for
propagating register copies.
To address this, the patch makes following changes to fwprop.c:
(a) Add support for handling UNSPEC in propagate_rtx_1 in a similar
manner to simplify_replace_fn_rtx.
(b) Allow propagating def inside a loop if source of def is a register
in forward_propagate_into.
AFAIU, replacing register by another register shouldn't increase cost.
(c) Integrate fwprop and fwprop_addr, and make fwprop_addr propagate
register copies.

With the patch, fwprop_addr propagates reg:SI 97 in insn 23 and deletes insn 22,
which eliminates the redundant mov.

Does this patch look OK ?
Bootstrapped + tested on x86_64-unknown-linux-gnu and aarch64-linux-gnu.
Cross-testing with SVE in progress.

Thanks,
Prathamesh