[Bug middle-end/94035] New: Wrong optimization: conditional equivalence vs. values with several representations

[ch3root at openwall dot com]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=94035

            Bug ID: 94035
           Summary: Wrong optimization: conditional equivalence vs. values
                    with several representations
           Product: gcc
           Version: 10.0
            Status: UNCONFIRMED
          Severity: normal
          Priority: P3
         Component: middle-end
          Assignee: unassigned at gcc dot gnu.org
          Reporter: ch3root at openwall dot com
  Target Milestone: ---

The problem happens when:
- conditional equivalence propagation replaces an expression with a variable or
a constant that has the same value but a different representation, and
- this happens in a computation where representation is accessed.

Example with a pseudo-denormal in long double:

----------------------------------------------------------------------
#include <stdio.h>

__attribute__((noipa,optnone)) // imagine it in a separate TU
static void *opaque(void *p) { return p; }

int main()
{
    long double x, y;

    unsigned char *px = (unsigned char *)&x;
    unsigned char *py = (unsigned char *)&y;

    // make x pseudo-denormal
    x = 0;
    px[7] = 0x80;
    opaque(&x); // hide it from the optimizer

    y = x;

    if (y == 0x1p-16382l)
        printf("py[8] = %d\n", py[8]);
    printf("py[8] = %d\n", py[8]);
}
----------------------------------------------------------------------
$ gcc -std=c11 -pedantic -Wall -Wextra -Wno-attributes -O3 test.c && ./a.out
py[8] = 1
py[8] = 0
----------------------------------------------------------------------
gcc x86-64 version: gcc (GCC) 10.0.1 20200304 (experimental)
----------------------------------------------------------------------

The value 0x1p-16382l admits two representations:

00 00 80 00 00 00 00 00 00 00  pseudo-denormal
00 01 80 00 00 00 00 00 00 00  normalized value

So both 0 and 1 for py[8] are fine but the testcase should print the same value
both times, i.e. the representation of y should be stable.

DR 260 Q1 allows for unstable representation but IMO this is wrong.