On Wed, May 13, 2015 at 8:53 PM, Richard Henderson <r...@redhat.com> wrote:
> On 05/13/2015 11:11 AM, Uros Bizjak wrote:
>> We can use general_operand instead of some_operand.
>>
>> 2015-05-13 Uros Bizjak <ubiz...@gmail.com>
>>
>> * config/alpha/alpha.md (extendqidi2): Use general_operand
>> instead of some_operand for operand[1] predicate.
>> (extendhidi2): Ditto.
>> (cbranchdi4): Use general_operand instead of some_operand
>> for operand[1] and operands[2] predicates.
>> (cstoredi4): Ditto.
>> * config/alpha/predicates.md (some_operand): Remove unused predicate.
>> (some_ni_operand): Ditto.
>>
>> Tested on alpha-linux-gnu.
>>
>> Richard, does this look OK to you, or is there any other reason that
>> general_operand predicates were not used here?
>
> For the extensions, it was put in by Kenner in 1997 (90f6b60d), to improve
> code
> for unaligned memories. That code was removed in 2011 by me (8b2983a3), so I
> think dropping some_operand there is fine.
>
> For the conditionals, it was added in 2004 by me (62350d6c), and that code is
> still there. Specifically,
>
> @@ -3177,11 +3177,17 @@ alpha_emit_conditional_branch (enum rtx_code code)
> cmp_code = NIL, branch_code = code;
>
> /* If the constants doesn't fit into an immediate, but can
> be generated by lda/ldah, we adjust the argument and
> compare against zero, so we can use beq/bne directly. */
> - else if (GET_CODE (op1) == CONST_INT && (code == EQ || code == NE))
> + /* ??? Don't do this when comparing against symbols, otherwise
> + we'll reduce (&x == 0x1234) to (&x-0x1234 == 0), which will
> + be declared false out of hand (at least for non-weak). */
> + else if (GET_CODE (op1) == CONST_INT
> + && (code == EQ || code == NE)
> + && !(symbolic_operand (op0, VOIDmode)
> + || (GET_CODE (op0) == REG && REG_POINTER (op0))))
>
> If I didn't use some_operand, the SYMBOL_REF would be lowered and we'll only
> see a REG here. Searching the mail archive I find
>
> https://gcc.gnu.org/ml/gcc-patches/2004-02/msg02436.html
>
> pointing to the test case gcc.dg/20040123-1.c
>
> Perhaps debugging that testcase to see what's reaching a_e_c_b in these modern
> times will tell you what's most appropriate.
The reason for removing some_operand is, that it is effectively
exactly equal to general_operand:
from gensupport.c:
{"general_operand", false, true, {SUBREG, REG, MEM}},
the second "true" stands for "bool allows_const_p", where:
if (std_preds[i].allows_const_p)
for (j = 0; j < NUM_RTX_CODE; j++)
if (GET_RTX_CLASS (j) == RTX_CONST_OBJ)
add_predicate_code (pred, (enum rtx_code) j);
and from rtl.def, these are all RTX_CONST_OBJs:
rtl.def:DEF_RTL_EXPR(CONST_INT, "const_int", "w", RTX_CONST_OBJ)
rtl.def:DEF_RTL_EXPR(CONST_WIDE_INT, "const_wide_int", "", RTX_CONST_OBJ)
rtl.def:DEF_RTL_EXPR(CONST_FIXED, "const_fixed", "www", RTX_CONST_OBJ)
rtl.def:DEF_RTL_EXPR(CONST_DOUBLE, "const_double",
CONST_DOUBLE_FORMAT, RTX_CONST_OBJ)
rtl.def:DEF_RTL_EXPR(CONST_VECTOR, "const_vector", "E", RTX_CONST_OBJ)
rtl.def:DEF_RTL_EXPR(CONST, "const", "e", RTX_CONST_OBJ)
rtl.def:DEF_RTL_EXPR(LABEL_REF, "label_ref", "u", RTX_CONST_OBJ)
rtl.def:DEF_RTL_EXPR(SYMBOL_REF, "symbol_ref", "s0", RTX_CONST_OBJ)
rtl.def:DEF_RTL_EXPR(HIGH, "high", "e", RTX_CONST_OBJ)
This is exactly the list of constant operands (minus const_fixed) from
some_operand predicate:
(define_predicate "some_operand"
(ior (match_code "reg,mem,const_int,const_wide_int,const_double,const_vector,
label_ref,symbol_ref,const,high")
(and (match_code "subreg")
(match_test "some_operand (SUBREG_REG (op), VOIDmode)"))))
So, the only question is about SUBREG processing here, but I don't
think it matters here.
Uros.
