[Bug target/62173] [5.0 regression] [AArch64] Performance regression due to r213488

[jiwang at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=62173

--- Comment #8 from Jiong Wang <jiwang at gcc dot gnu.org> ---
while above associate virtual_stack_var_rtx with constant offset actually cause
another hidding issue.

I assume after the association, we are generating cleaner insn sequences.

given the following testcase

void bar(int i)              
{                        
  char A[10];
  char B[10];     
  char C[10];       
  g(A);                        
  g(B);
  g(C);                              
  f(A[i]);                 
  f(B[i]);                        
  f(C[i]);                   
  return;               
} 

it will generate the following assembly

bar:
    stp    x29, x30, [sp, -96]!
    add    x29, sp, 0
    stp    x21, x22, [sp, 32]
    add    x22, x29, 48
    add    x21, x29, 64
    stp    x19, x20, [sp, 16]
    sxtw    x19, w0
    mov    x0, x22
    add    x20, x29, 80
    bl    g
    mov    x0, x21
    bl    g
    mov    x0, x20
    bl    g
    ldrb    w0, [x22, x19]
    bl    f
    ldrb    w0, [x21, x19]
    bl    f
    ldrb    w0, [x20, x19]
    bl    f
    ldp    x19, x20, [sp, 16]
    ldp    x21, x22, [sp, 32]
    ldp    x29, x30, [sp], 96
    ret

above assembly is bad, because extra callee saved registers are used which we
should rematerialize those frame address instead of keeping them in callee
saved to avoid extra load/store.

looks like current register allocation don't rematerialize above situation.