[Bug target/23831] [4.1 Regression] ICE in immed_double_const with vectorized multipication

[uros at kss-loka dot si]

------- Additional Comments From uros at kss-loka dot si  2005-09-15 11:53 -------
He problem is in following RTL:

(insn 30 29 31 1 (set (reg:V4SI 75)
        (mem/u/i:V4SI (symbol_ref/u:SI ("*.LC0") [flags 0x2]) [3 S16 A128])) 
541 {*movv4si_internal} (nil)
    (expr_list:REG_EQUAL (const_vector:V4SI [
                (const_int 5 [0x5])
                (const_int 5 [0x5])
                (const_int 5 [0x5])
                (const_int 5 [0x5])
            ])
        (nil)))

...

(insn 33 32 34 1 (set (subreg:TI (reg:V4SI 78) 0)
        (lshiftrt:TI (subreg:TI (reg:V4SI 75) 0)
            (const_int 32 [0x20]))) 709 {sse2_lshrti3} (nil)
    (nil))

Somehow a TImode constant is produced, and it triggers the check below:

      gcc_assert (GET_MODE_CLASS (mode) == MODE_INT
		  || GET_MODE_CLASS (mode) == MODE_PARTIAL_INT
		  /* We can get a 0 for an error mark.  */
		  || GET_MODE_CLASS (mode) == MODE_VECTOR_INT
		  || GET_MODE_CLASS (mode) == MODE_VECTOR_FLOAT);

      /* We clear out all bits that don't belong in MODE, unless they and
	 our sign bit are all one.  So we get either a reasonable negative
	 value or a reasonable unsigned value for this mode.  */
      width = GET_MODE_BITSIZE (mode);
      if (width < HOST_BITS_PER_WIDE_INT
	  && ((i0 & ((HOST_WIDE_INT) (-1) << (width - 1)))
	      != ((HOST_WIDE_INT) (-1) << (width - 1))))
	i0 &= ((HOST_WIDE_INT) 1 << width) - 1, i1 = 0;
      else if (width == HOST_BITS_PER_WIDE_INT
	       && ! (i1 == ~0 && i0 < 0))
	i1 = 0;
      else
	/* We should be able to represent this value as a constant.  */
>>>	gcc_assert (width <= 2 * HOST_BITS_PER_WIDE_INT);

HOSTS_BITS_PER_WIDE_INT is 32 bit for ia32, and TImode constants are 128 bit 
wide. I don't know the correct solution here. If the assert is commented out, 
the ICE goes away and produced code works as expected.

-- 


http://gcc.gnu.org/bugzilla/show_bug.cgi?id=23831