Using shifts by 1 on K6,K7 and i386

Using shifts by 1 on K6,K7 and i386

[Martin Erhardsen]

Why doesn't gcc use logical shifts, arithmetric shifts and rotates by 1 
instead of immediate operands on the K6,K7 and i386.
It saves a byte and has the same latency (except on i486, 
where it uses an extra cycle). I have measured this on 
my K6-2 and looked it up for K7 and i386
I propose that the current predicate for using shifts 
and rotates by 1

(TARGET_PENTIUM || TARGET_PENTIUMPRO)

be replaced by this

(!TARGET_I486 || optimize_size)

Re: Using shifts by 1 on K6,K7 and i386

[Jeffrey A Law]

  In message < 000701bfe33f$1e751e60$b495f9c3@mincomputer >you write:
  > Why doesn't gcc use logical shifts, arithmetric shifts and rotates by 1 
  > instead of immediate operands on the K6,K7 and i386.
Because nobody's told us what is best for this case on the K6, K7 and
i386.

  > It saves a byte and has the same latency (except on i486, 
  > where it uses an extra cycle). I have measured this on 
  > my K6-2 and looked it up for K7 and i386
OK.

  > I propose that the current predicate for using shifts 
  > and rotates by 1
So, submit a patch which changes i386.md appropriately.

THanks,
jeff

Re: Using shifts by 1 on K6,K7 and i386

[Martin Erhardsen]

----- Original Message ----- 
From: "Jeffrey A Law" <law@cygnus.com>
To: "Martin Erhardsen" <MartinErhardsen@mail.tele.dk>
Cc: <gcc@gcc.gnu.org>
Sent: Thursday, July 06, 2000 3:06 AM
Subject: Re: Using shifts by 1 on K6,K7 and i386 


> 
>   In message < 000701bfe33f$1e751e60$b495f9c3@mincomputer >you write:
>   > Why doesn't gcc use logical shifts, arithmetric shifts and rotates by 1 
>   > instead of immediate operands on the K6,K7 and i386.
> Because nobody's told us what is best for this case on the K6, K7 and
> i386.
> 
>   > It saves a byte and has the same latency (except on i486, 
>   > where it uses an extra cycle). I have measured this on 
>   > my K6-2 and looked it up for K7 and i386
> OK.
> 
>   > I propose that the current predicate for using shifts 
>   > and rotates by 1
> So, submit a patch which changes i386.md appropriately.

Okay, I'll check the documentation for the intel processors 
to make sure that this is the right thing to do for those processors 
too, and then I'll start writing my patch.

Re: Using shifts by 1 on K6,K7 and i386

[Richard Henderson]

On Sat, Jul 01, 2000 at 11:30:56AM +0200, Martin Erhardsen wrote:
> I propose that the current predicate for using shifts 
> and rotates by 1
> 
> (TARGET_PENTIUM || TARGET_PENTIUMPRO)
> 
> be replaced by this
> 
> (!TARGET_I486 || optimize_size)

Actaully, if any change is made, it should be of the form
`TARGET_USE_SHIFT1', much like the existing TARGET_USE_LEAVE
and friends.


r~

Re: Using shifts by 1 on K6,K7 and i386

[Martin Erhardsen]

----- Original Message ----- 
From: "Richard Henderson" <rth@cygnus.com>
To: "Martin Erhardsen" <MartinErhardsen@mail.tele.dk>
Cc: <gcc@gcc.gnu.org>
Sent: Friday, July 07, 2000 2:24 AM
Subject: Re: Using shifts by 1 on K6,K7 and i386


> On Sat, Jul 01, 2000 at 11:30:56AM +0200, Martin Erhardsen wrote:
> > I propose that the current predicate for using shifts 
> > and rotates by 1
> > 
> > (TARGET_PENTIUM || TARGET_PENTIUMPRO)
> > 
> > be replaced by this
> > 
> > (!TARGET_I486 || optimize_size)
> 
> Actaully, if any change is made, it should be of the form
> `TARGET_USE_SHIFT1', much like the existing TARGET_USE_LEAVE
> and friends.
> 

I have looked at the 486 manual, and it is not that simple. Shifts by one are 
slower only when using register operands on the 486. This means that the 
ideally the predicate should be something like this

(!(TARGET_I486 && register_operand(operands[0],SImode)) || optimize_size)

I'm not sure what the second operand to memory_operand does and if it is 
necessary, but I would need at least one operand to a TARGET_USE_SHIFT1
macro.

Re: Using shifts by 1 on K6,K7 and i386

[Jeffrey A Law]

  In message < 006c01bfe813$524c0680$3595f9c3@mincomputer >you write:
  > > Actaully, if any change is made, it should be of the form
  > > `TARGET_USE_SHIFT1', much like the existing TARGET_USE_LEAVE
  > > and friends.
Richard is correct here.

  > I have looked at the 486 manual, and it is not that simple. Shifts by one
  > are slower only when using register operands on the 486. This means that
  > the ideally the predicate should be something like this
  > 
  > (!(TARGET_I486 && register_operand(operands[0],SImode)) || optimize_size)
  > 
  > I'm not sure what the second operand to memory_operand does and if it is 
  > necessary, but I would need at least one operand to a TARGET_USE_SHIFT1
  > macro.
The second operand to "register_operand" is the mode of the operand.  In this
case it should be the mode of the value that you want to shift.

Instead of using SImode, just use GET_MODE (operands[0]) for the second
argument.

Yes, you'll need to make the TARGET_USE_SHIFT1 change like Richard suggested
too -- and if you need an argument, that's not a problem either.  There's no
reason that the new TARGET_USE_SHIFT1 couldn't accept an argument that I'm
aware of.


jeff