On Wed, Nov 20, 2013 at 5:11 PM, Ilya Enkovich <[email protected]> wrote:
> 2013/11/20 Uros Bizjak <[email protected]>:
>> On Wed, Nov 20, 2013 at 1:32 PM, Ilya Enkovich <[email protected]>
>> wrote:
>>
>>>> >> > Here is a patch to add size relocation and instruction to obtain
>>>> >> > object's size in i386 target.
>>>> >>
>>>> >> +(define_insn "move_size_reloc_<mode>"
>>>> >> + [(set (match_operand:SWI48 0 "register_operand" "=r")
>>>> >> + (match_operand:<MODE> 1 "size_relocation" "Z"))]
>>>> >> + ""
>>>> >> +{
>>>> >> + return "mov{<imodesuffix>}\t{%1, %0|%0, %1}";
>>>> >>
>>>> >> Please don't change x86_64_immediate_operand just to use "Z"
>>>> >> constraint The predicate is used in a couple of other places that for
>>>> >> sure don't accept your change.
>>>> >>
>>>> >> Better write this insn in an explicit way (see for example
>>>> >> tls_initial_exec_64_sun). Something like:
>>>> >>
>>>> >> (define_insn "move_size_reloc_<mode>"
>>>> >> [(set (match_operand:SWI48 0 "register_operand" "=r")
>>>> >> (unspec:SWI48
>>>> >> [(match_operand 1 "symbolic_operand" "..." )]
>>>> >> UNSPEC_SIZEOF))]
>>>> >> ""
>>>> >> "mov{<imodesuffix>}\t{%a1@SIZE, %0|%0, %a1@SIZE}")
>>>> >>
>>>> >> You will probably need to define new operand 1 predicate and constraint.
>>>> >>
>>>> >> Uros.
>>>> >
>>>> > Hi, Uros! Thanks for comments! Here is what I got trying to follow
>>>> > your suggestion. Does it look better?
>>>>
>>>> You actually don't need any operand modifiers in the insn template. Simply
>>>> use:
>>>>
>>>> "mov{<imodesuffix>}\t{%1@SIZE, %0|%0, %1@SIZE}"
>>>>
>>>> and you will automatically get
>>>>
>>>> "movl $zzz, %eax" or "mov %eax, OFFSET FLAT: zzz".
>>>>
>>>> Since your pattern allows only symbolic_operand, there is no reload,
>>>> so you can avoid the constraint alltogether.
>>>>
>>>> BTW: Did you consider various -mcmodel=... options? For DImode moves,
>>>> you should check x86_64_zext_immediate_operand predicate and output
>>>> either "movl $zzz, %eax" or "movabs $zzz, %rax". There is no movq with
>>>> 64bit immediate. Please see movdi pattern.
>>>
>>> Yep, for large objects it may work wrongly. Does anyone use static objects
>>> >4Gb? :)
>>>
>>> Large address does not mean large object but seems we have to be
>>> conservative here. I added x86_64_zext_immediate_operand check with
>>> additional CM_KERNEL check because in this model object size should always
>>> fit 32 bits.
>>
>> IMO, we should list code models that support object sizes > 31bits for
>> 64bit target. The object size in small models will never be > 31bits
>> (and never negative), so we can use movl unconditionally.
>
> For medium models x86_64_zext_immediate_operand returns true for
> object is known to go to lower 2Gb space. It should allow us to use
> movl. Why do you always emit movabs for medium model?
CM_MEDIUM has unlimited data size.
i386-opts.h: CM_MEDIUM, /* Assumes code fits in the low 31
bits; data unlimited. */
The x86_64_zext_immediate_operand allows _address_ to be loaded by
movl. The @SIZE reloc returns object size, which is unlimited and has
no connection to its address. For CM_MEDIUM,
x86_64_zext_immediate_operand allows:
return (ix86_cmodel == CM_SMALL
|| (ix86_cmodel == CM_MEDIUM
&& !SYMBOL_REF_FAR_ADDR_P (op)));
Uros.
