[Bug target/77822] [6 Regression] arm64 Error: immediate value out of range 0 to 63 at operand 3

[ktkachov at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=77822

--- Comment #24 from ktkachov at gcc dot gnu.org ---
(In reply to Dominik Vogt from comment #23)
> Regarding the ARM patch:
> 
> +  {
> +    if (!IN_RANGE (INTVAL (operands[2]) + INTVAL (operands[3]),
> +		   1, GET_MODE_BITSIZE (DImode) - 1))
> +     FAIL;
> +  }
> 
> Isn't this patch too simple?  On s390x we get "pos" -N (operand 3) with
> "size" M, so the correct check should be:
> 
>   if (! IN_RANGE (INTVAL (operands[3]), 0, GET_MODE_BITSIZE (DImode) - 1)
>       || IN_RANGE (INTVAL (operands[3]) + INTVAL (operands[2]),
>                    1, GET_MODE_BITSIZE (DImode)))
>     FAIL;
> 
> It's just luck that the test program does not generate something like
> 
>   (zero_extract (...) (-1) (16))
> 
> Which can only be caught by testing "pos" individually.
> 
> --
> 
> Also, I don't understand why the valid range is "1, GET_MODE_BITSIZE
> (DImode) - 1", shouldn't it be "1, GET_MODE_BITSIZE (DImode)" without the "-
> 1"?

operands[3] also gets the aarch64_simd_shift_imm_di predicate which checks that
IN_RANGE (INTVAL (op), 0, 63), so we should be okay