[Bug tree-optimization/114682] New: const_array[i].b where i is PHI<0,1,2> is not folded

[Bug tree-optimization/114682] New: const_array[i].b where i is PHI<0,1,2> is not folded

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114682

            Bug ID: 114682
           Summary: const_array[i].b where i is PHI<0,1,2> is not folded
           Product: gcc
           Version: 14.0
            Status: UNCONFIRMED
          Keywords: missed-optimization
          Severity: enhancement
          Priority: P3
         Component: tree-optimization
          Assignee: unassigned at gcc dot gnu.org
          Reporter: pinskia at gcc dot gnu.org
  Target Milestone: ---

Take:
```
#include <cstddef>

typedef struct cts_mode_name2id_st {
    unsigned int id;
    const char *name;
} CTS_MODE_NAME2ID;


/* The value assigned to 0 is the default */
#define CTS_CS1 0
#define CTS_CS2 1
#define CTS_CS3 2


/* OSSL_CIPHER_PARAM_CTS_MODE Values */
# define OSSL_CIPHER_CTS_MODE_CS1 "CS1"
# define OSSL_CIPHER_CTS_MODE_CS2 "CS2"
# define OSSL_CIPHER_CTS_MODE_CS3 "CS3"

static  CTS_MODE_NAME2ID cts_modes[] =
{
    { CTS_CS1, OSSL_CIPHER_CTS_MODE_CS1 },
    { CTS_CS2, OSSL_CIPHER_CTS_MODE_CS2 },
    { CTS_CS3, OSSL_CIPHER_CTS_MODE_CS3 },
};

#define OSSL_NELEM(a) (sizeof(a)/sizeof(a[0]))
int OPENSSL_strcasecmp(const char*, const char*);

int ossl_cipher_cbc_cts_mode_name2id(const char *name)
{
    size_t i;

    for (i = 0; i < OSSL_NELEM(cts_modes); ++i) {
        if (OPENSSL_strcasecmp(name, cts_modes[i].name) == 0)
            return (int)cts_modes[i].id;
    }
    return -1;
}

```

We end up with the following in the .optimized at -O2 or -O3:
```
  # i_3 = PHI <2(4), 0(2), 1(3)>
  _2 = cts_modes[i_3].id;
  _12 = (int) _2;
```

But cts_modes[i_3].id here is could be constant folded by the compiler into the
same constants as i really.

I suspect it is because we don't have a PRE post unrolling pass which could
handle that.

This shows up in openssl
https://github.com/openssl/openssl/blob/master/providers/implementations/ciphers/cipher_cts.c#L55

[Bug tree-optimization/114682] const_array[i].b where i is PHI<0,1,2> is not folded

[pinskia at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114682

--- Comment #1 from Andrew Pinski <pinskia at gcc dot gnu.org> ---
Note I was imspired to look into this because of
https://github.com/llvm/llvm-project/issues/88274 (note LLVM does a worse job
here and does not constant loads the string either).

[Bug tree-optimization/114682] const_array[i].b where i is PHI<0,1,2> is not folded

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114682

Richard Biener <rguenth at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
             Status|UNCONFIRMED                 |NEW
     Ever confirmed|0                           |1
   Last reconfirmed|                            |2024-04-11

--- Comment #2 from Richard Biener <rguenth at gcc dot gnu.org> ---
Note it isn't necessarily always an improvement.  It's also not really PRE
but FRE when you do PHI translation.  There's an argument that we should
try PHI translation in FRE to at least catch cases like this where
the replacement is simply

  # i_3 = PHI <2(4), 0(2), 1(3)>
  _12 = (int) i_3;

Note there's also that "frankenstein" phiprop pass which does remotely
similar things.