[Bug tree-optimization/41089] [4.5 Regression] r147980 (New SRA) breaks stdargs

["ubizjak at gmail dot com" <gcc-bugzilla at gcc dot gnu dot org>]

------- Comment #20 from ubizjak at gmail dot com  2009-09-10 17:29 -------
I have compared the dumps of two functions:

--function 1, WORKS--
extern void abort (void);

int foo_arg;
long x;

void
f4 (int i, ...)
{
  __builtin_va_list ap;

  __builtin_va_start(ap,i);
  x = __builtin_va_arg(ap,double);
  foo_arg = __builtin_va_arg(ap,int);
  __builtin_va_end(ap);
}

--stdarg dump of function 1--
;; Function f4 (f4)

f4: va_list escapes 0, needs to save 16 GPR units and 3 FPR units.
f4 (int i)
{
  struct  ap;
  int foo_arg.2;
  int D.2032;
  int * D.2031;
  long unsigned int D.2030;
  int * D.2029;
  long int D.2028;
  long int x.1;
  double D.2026;
  int D.2025;
  double * D.2024;
  long int D.2022;
  long unsigned int iftmp.0;
  double * D.2018;
  long int D.2017;
  int D.2016;
  void * D.2015;

<bb 2>:
  __builtin_va_start (&ap, 0);
  D.2015_2 = ap.__base;
  D.2016_3 = ap.__offset;
  D.2017_4 = (long int) D.2016_3;
  D.2018_5 = (double *) D.2015_2;
  if (D.2017_4 <= 47)
    goto <bb 3>;
  else
    goto <bb 4>;

<bb 3>:
  D.2022_6 = D.2017_4 + -48;
  iftmp.0_7 = (long unsigned int) D.2022_6;
  goto <bb 5>;

<bb 4>:
  iftmp.0_8 = (long unsigned int) D.2017_4;

<bb 5>:
  # iftmp.0_1 = PHI <iftmp.0_7(3), iftmp.0_8(4)>
  D.2024_9 = D.2018_5 + iftmp.0_1;
  D.2017_10 = D.2017_4 + 8;
  D.2025_11 = (int) D.2017_10;
  ap.__offset = D.2025_11;
  D.2026_12 = *D.2024_9;
  x.1_13 = (long int) D.2026_12;
  x = x.1_13;
  D.2028_16 = (long int) D.2025_11;
  D.2029_17 = (int *) D.2015_2;
  D.2030_18 = (long unsigned int) D.2028_16;
  D.2031_19 = D.2029_17 + D.2030_18;
  D.2028_20 = D.2028_16 + 8;
  D.2032_21 = (int) D.2028_20;
  ap.__offset = D.2032_21;
  foo_arg.2_22 = *D.2031_19;
  foo_arg = foo_arg.2_22;
  __builtin_va_end (&ap);
  return;

}

--function 2, FAILS--
extern void abort (void);

int foo_arg;
long x;

static void
foo (int v, __builtin_va_list ap)
{
  switch (v)
    {
    case 5: foo_arg = __builtin_va_arg(ap,int); break;
    default: abort ();
    }
}

void
f4 (int i, ...)
{
  __builtin_va_list ap;

  __builtin_va_start(ap,i);
  x = __builtin_va_arg(ap,double);
  foo (i, ap);
  __builtin_va_end(ap);
}

--stdarg dump of function 2--
;; Function f4 (f4)

f4: va_list escapes 0, needs to save 8 GPR units and 3 FPR units.
f4 (int i)
{
  long int D.2051;
  int * D.2050;
  long unsigned int D.2049;
  int * D.2048;
  int foo_arg.2;
  struct  ap;
  long int x.1;
  double D.2033;
  int D.2032;
  double * D.2031;
  long int D.2029;
  long unsigned int iftmp.0;
  double * D.2025;
  long int D.2024;
  int D.2023;
  void * D.2022;

<bb 2>:
  __builtin_va_start (&ap, 0);
  D.2022_2 = ap.__base;
  D.2023_3 = ap.__offset;
  D.2024_4 = (long int) D.2023_3;
  D.2025_5 = (double *) D.2022_2;
  if (D.2024_4 <= 47)
    goto <bb 3>;
  else
    goto <bb 4>;

<bb 3>:
  D.2029_6 = D.2024_4 + -48;
  iftmp.0_7 = (long unsigned int) D.2029_6;
  goto <bb 5>;

<bb 4>:
  iftmp.0_8 = (long unsigned int) D.2024_4;

<bb 5>:
  # iftmp.0_1 = PHI <iftmp.0_7(3), iftmp.0_8(4)>
  D.2031_9 = D.2025_5 + iftmp.0_1;
  D.2024_10 = D.2024_4 + 8;
  D.2032_11 = (int) D.2024_10;
  ap.__offset = D.2032_11;
  D.2033_12 = *D.2031_9;
  x.1_13 = (long int) D.2033_12;
  x = x.1_13;
  switch (i_14(D)) <default: <L4>, case 5: <L3>>

<L3>:
  D.2051_23 = (long int) D.2032_11;
  D.2050_24 = (int *) D.2022_2;
  D.2049_25 = (long unsigned int) D.2051_23;
  D.2048_26 = D.2050_24 + D.2049_25;
  foo_arg.2_27 = *D.2048_26;
  foo_arg = foo_arg.2_27;
  __builtin_va_end (&ap);
  return;

<L4>:
  abort ();

}


Tracing through the tree-stdarg.c, the difference is in the number of calls to
va_list_counter_struct_op from this place:

                {
                  if (get_gimple_rhs_class (gimple_assign_rhs_code (stmt))
                      == GIMPLE_SINGLE_RHS)
                    {
                      /* Check for ap[0].field = temp.  */
>>>		      if (va_list_counter_struct_op (&si, lhs, rhs, true))
                        continue;

                      /* Check for temp = ap[0].field.  */
                      else if (va_list_counter_struct_op (&si, rhs, lhs,
                                                          false))
                        continue;
                    }

So indeed, in the failing case, there is only one assignment to  ap._offset and
in the working case, there are two assignments, resulting in correct number of
GPR bytes (== 2 registers) saved.


-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=41089