Hi,
The following test:

#include "arm_sve.h"

svint8_t
test_s8(int8_t *x)
{
  return svld1rq_s8 (svptrue_b8 (), &x[0]);
}

ICE's with -march=armv8.2-a+sve -O1 -fno-tree-ccp -fno-tree-forwprop:
during GIMPLE pass: fre
pr107920.c: In function ‘test_s8’:
pr107920.c:7:1: internal compiler error: in execute_todo, at passes.cc:2140
    7 | }
      | ^
0x7b03d0 execute_todo
        ../../gcc/gcc/passes.cc:2140

because of incorrect handling of virtual operands in svld1rq_impl::fold:
 # VUSE <.MEM>
  _5 = MEM <vector(16) signed char> [(signed char * {ref-all})x_3(D)];
  _4 = VEC_PERM_EXPR <_5, _5, { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
12, 13, 14, 15, ... }>;
  # VUSE <.MEM_2(D)>
  return _4;

The attached patch tries to fix the issue by building the replacement
statements in gimple_seq, and passing it to gsi_replace_with_seq_vops,
which resolves the ICE, and results in:
  <bb 2> :
  # VUSE <.MEM_2(D)>
  _5 = MEM <vector(16) signed char> [(signed char * {ref-all})x_3(D)];
  _4 = VEC_PERM_EXPR <_5, _5, { 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11,
12, 13, 14, 15, ... }>;
  # VUSE <.MEM_2(D)>
  return _4;

Bootstrapped+tested on aarch64-linux-gnu.
OK to commit ?

Thanks,
Prathamesh