Re: Back end question.

["Dmitry" <mittie at ukr dot net>]

I know that const_int isn't smart enough, but what if I use if_then_else rtx, and test there a register against zero for example, and set in result a CCCZ mode accordingly?
I mean such a pattern:
(set (reg:CCZ ST0) (if_then_else (eq:HI (match_dup 0) (const_int 0)) (const_int 1) (const_int 0))
And in a single CC mode I must use a single set rtx which sets a back-end's defined single value to the status reg?
Thanks.
Dmitry.

--- Исходное сообщение ---
 От кого: "Ian Lance Taylor" <iant@google.com> 
 Кому: "Dmitry" <mittie@ukr.net> 
 Дата: 6 мая 2011, 18:54:52 
 Тема: Re: Back end question. 
 


> "Dmitry" <mittie@ukr.net> writes:
> 
> > So if for example I could write something like this
> > (set (reg:CCCZ ST0) (const_int 1))
> > And the compiler would think that by that I've set a carry and zero flag for example?
> > And how later a branch instruction could determine whether a carry or a zero flag was previously set in that CCCZ mode?
> 
> Yes, you could write that, though it would be unlikely to succeed. The
> compiler doesn't really know anything about carry or zero flags. It
> only knows what your backend says. The branch instruction in your
> backend could look for (const_int 1). The problem with this example is
> that you haven't mentioned the value being compared, so the branch
> instruction could match any test. But if you do write an expression
> which uses the registers, then you could write pretty much any
> expression, as long as some insn sets it and the branch instruction
> tests it.
> 
> Ian
> 
> 
> > Thanks.
> > Dmitry.
> >
> > --- Исходное сообщение ---
> > От кого: "Ian Lance Taylor" <iant@google.com> 
> > Кому: "Dmitry" <mittie@ukr.net> 
> > Дата: 6 мая 2011, 18:28:46 
> > Тема: Re: Back end question. 
> > 
> >
> >
> >> "Dmitry" <mittie@ukr.net> writes:
> >> 
> >> > Thanks Ian for your patience. But still I get that I can set several bits with one compare, but how can I use one compare if I need to compare with two different values?
> >> > To set a zero bit I should compare with 0 to set an overflow bit I should compare with 65535?
> >> 
> >> I think you are approaching this the wrong way. The value that you set
> >> in the flags register is not the expression which the flags register
> >> stores. It is the expression which the flag register is about. That
> >> expression then has to match what the branch instruction looks for.
> >> 
> >> It is not necessary to compare with 65535 to set the overflow flag.
> >> It's just necessary to have some comparison which the branch instruction
> >> will match.
> >> 
> >> In any case it is quite difficult to write C code which does a single
> >> comparison and then does two branches, one on the overflow flag and one
> >> on the zero flag.
> >> 
> >> Ian
> >> 
> >> > --- Исходное сообщение ---
> >> > От кого: "Ian Lance Taylor" <iant@google.com> 
> >> > Кому: "Dmitry" <mittie@ukr.net> 
> >> > Дата: 5 мая 2011, 21:06:29 
> >> > Тема: Re: Back end question. 
> >> > 
> >> >
> >> >
> >> >> "Dmitry" <mittie@ukr.net> writes:
> >> >> 
> >> >> > Yes Ian I understand the whole idea of different CC modes but, what I was asking is how could I load actual values to the separate status bits. I mean in the i386 example you gave "(set (reg:CCC FLAGS_REG) (compare:CCC ..." it sets the only one bit with one set rtx. But if it had CCCZ mode for example (carry + zero bits) it should go like:
> >> >> > (set (reg:CCCZ FLAGS_REG) (compare:CCCZ ... - carry bit computation
> >> >> > (set (reg:CCCZ FLAGS_REG) (compare:CCCZ ... - zero bit computation
> >> >> > or a parallel of some sort?
> >> >> > So the question is how could I set the correct values for carry and zero bits? What should go after "(set (reg:CCCZ FLAGS_REG) ..." for two separate status bits?
> >> >> 
> >> >> A set to FLAGS_REG of a compare does not set one bit. It sets whatever
> >> >> bits are represented by the mode. So you don't need multiple sets of a
> >> >> single register. That doesn't even make sense in gcc's RTL
> >> >> representation.
> >> >> 
> >> >> Think about how the value is used: it is used in a conditional
> >> >> instruction which is looking at the compare. That conditional can test
> >> >> whichever bits the mode permits.
> >> >> 
> >> >> Ian
> >> >> 
> >> >> > --- Исходное сообщение ---
> >> >> > От кого: "Ian Lance Taylor" <iant@google.com> 
> >> >> > Кому: "Dmitry" <mittie@ukr.net> 
> >> >> > Дата: 5 мая 2011, 19:44:55 
> >> >> > Тема: Re: Back end question. 
> >> >> > 
> >> >> >
> >> >> >
> >> >> >> "Dmitry" <mittie@ukr.net> writes:
> >> >> >> 
> >> >> >> > Yes but as you've said, it's a carry flag setter, but how could this example be transformed if I need to store not only carry flag but also for example a zero flag? Could you write such an example, please?
> >> >> >> 
> >> >> >> You do exactly the same thing, but you define a mode which means that
> >> >> >> both the zero flag and the carry flag are valid. Then your test can
> >> >> >> look for that mode in the flags register. Don't look at this insn in
> >> >> >> isolation--look at it in conjunction with some conditional insn that
> >> >> >> tests the value in the flags register. The conditional test is going to
> >> >> >> be testing the result of the plus operation compared with zero, in some
> >> >> >> mode.
> >> >> >> 
> >> >> >> The i386 backend uses CCmode to represent all the flags, and uses the
> >> >> >> other modes to represent subsets of the flags.
> >> >> >> 
> >> >> >> Ian
> >> >> >> 
> >> >> >> > --- Исходное сообщение ---
> >> >> >> > От кого: "Ian Lance Taylor" <iant@google.com> 
> >> >> >> > Кому: "Dmitry" <mittie@ukr.net> 
> >> >> >> > Дата: 5 мая 2011, 17:41:38 
> >> >> >> > Тема: Re: Back end question. 
> >> >> >> > 
> >> >> >> >
> >> >> >> >
> >> >> >> >> "Dmitry" <mittie@ukr.net> writes:
> >> >> >> >> 
> >> >> >> >> > Ok but if people do not set each status bit individually, how do they set several not related bits in one CC mode and in one set rtx?
> >> >> >> >> 
> >> >> >> >> They represent the collection of bits as a mode, as you've been
> >> >> >> >> discussing.
> >> >> >> >> 
> >> >> >> >> E.g., from config/i386/i386.md
> >> >> >> >> 
> >> >> >> >> (define_insn "*<plusminus_insn><mode>3_cc_overflow"
> >> >> >> >> [(set (reg:CCC FLAGS_REG)
> >> >> >> >> (compare:CCC
> >> >> >> >> (plusminus:SWI
> >> >> >> >> (match_operand:SWI 1 "nonimmediate_operand" "<comm>0,0")
> >> >> >> >> (match_operand:SWI 2 "<general_operand>" "<r><i>,<r>m"))
> >> >> >> >> (match_dup 1)))
> >> >> >> >> (set (match_operand:SWI 0 "nonimmediate_operand" "=<r>m,<r>")
> >> >> >> >> (plusminus:SWI (match_dup 1) (match_dup 2)))]
> >> >> >> >> "ix86_binary_operator_ok (<CODE>, <MODE>mode, operands)"
> >> >> >> >> "<plusminus_mnemonic>{<imodesuffix>}\t{%2, %0|%0, %2}"
> >> >> >> >> [(set_attr "type" "alu")
> >> >> >> >> (set_attr "mode" "<MODE>")])
> >> >> >> >> 
> >> >> >> >> 
> >> >> >> >> Here the CCC mode (defined in i386-modes.def) means that the carry flag
> >> >> >> >> is valid.
> >> >> >> >> 
> >> >> >> >> Ian
> >> >> >> >> 
> >> >> >> >> 
> >> >> >> 
> >> >> >> 
> >> >> 
> >> >> 
> >> 
> >> 
> 
>