This is the mail archive of the
gcc-patches@gcc.gnu.org
mailing list for the GCC project.
Re: Patch: Consistently generate widening multiplies
On 04/07/2010 12:42 PM, Richard Guenther wrote:
> On Wed, Apr 7, 2010 at 2:40 PM, Bernd Schmidt <bernds@codesourcery.com> wrote:
>>> As I said earlier, please add a hunk to tree-cfg.c
>>> verify_gimple_assign_binary, the verifier is essentially documentation
>>> of gimple.
>>
>> Will do.
>>
>>>
>>> Further comments inline ...
>>
>> Thanks for reviewing.
>>
>>>> + if (rhs1_stmt != NULL && gimple_bb (rhs1_stmt) != gimple_bb (stmt))
>>>> + {
>>>> + gimple_stmt_iterator other_gsi = gsi_for_stmt (rhs1_stmt);
>>>> + gimple tmp_stmt;
>>>> + tree tmpvar1;
>>>> +
>>>> + tmpvar1 = create_tmp_var (TREE_TYPE (rhs1_convop), "wmulsrc");
>>>> + add_referenced_var (tmpvar1);
>>>> + tmpvar1 = make_ssa_name (tmpvar1, NULL);
>>>> +
>>>> + tmp_stmt = gimple_build_assign (tmpvar1, rhs1_convop);
>>>> + gsi_insert_before (&other_gsi, tmp_stmt, GSI_SAME_STMT);
>>>> + rhs1_convop = tmpvar1;
>>>> + }
>>>
>>> What's this code for?
>>
>> I was slightly worried that if I just used the argument of the
>> CONVERT_EXPR, using it in the multiply might give me a different value
>> in the presence of loops and conditions. So I insert a copy in the same
>> place as the conversion which I know will give me the correct value.
>> Something like the following is what I'm worried about:
>>
>> int a = 5, b = 3;
>> short xyz;
>> for (i = 0; i < 5; i++)
>> {
>> xyz = foo ();
>> if (i == 3)
>> a = (int)xyz;
>> }
>> ...
>> c = a * b;
>>
>> That turns into
>>
>> int a = 5, b = 3;
>> short xyz, copy;
>> for (i = 0; i < 5; i++)
>> {
>> xyz = foo ();
>> if (i == 3)
>> {
>> copy = xyz;
>> a = (int)xyz;
>> }
>> }
>> ...
>> c = (int)copy * b;
>>
>> rather than using xyz in the multiply.
>
> Ah, but you are in SSA form, so there's nothing bad that can happen.
> Thus you can simply delete the two blocks.
Okay, looking at some tree dumps, it seems we would never see this case
currently since the "a" that's multiplied isn't the result of the
extension but the result of a PHI. Ok. If we wanted to detect this
case, only then would we have to create copies and make a new PHI - correct?
Bernd