Change codegen for literals to emit xor when literal is zero.

Change codegen for literals to emit xor when literal is zero.

BUG=
TEST=

Committed to bleeding edge in r6104,

[Martin Maly, 2010-12-21 01:56:39]

Emit literal zero as xor is used in few places but while browsing through the
code I found that it is not so everywhere.

I ran the benchmarks (crypto, deltablue and friends) with and without my change
and detected no performance regression.

Thank you!
Martin

[William Hesse, 2010-12-21 07:22:14]

It is very important to make sure that Set() is not used in any place where we
want to preserve the flags - make sure that no new uses of Set() can occur
between the test and the branch in generated code.

I don't see the need for introducing a new Set(Register, int32_t).
The one place where it is used, you can simply wrap the int32 in an Immediate().


On 2010/12/21 01:56:39, Martin Maly wrote:
> Emit literal zero as xor is used in few places but while browsing through the
> code I found that it is not so everywhere.
> 
> I ran the benchmarks (crypto, deltablue and friends) with and without my
change
> and detected no performance regression.
> 
> Thank you!
> Martin

[mmaly, 2010-12-22 03:34:17]

Thanks William,

I removed the unneeded method and tracked down the conditional branches to
confirm that the new Set doesn't affect any of them. More details in separate
email.

Thank you!
Martin

On 2010/12/21 07:22:14, William Hesse wrote:
> It is very important to make sure that Set() is not used in any place where we
> want to preserve the flags - make sure that no new uses of Set() can occur
> between the test and the branch in generated code.
> 
> I don't see the need for introducing a new Set(Register, int32_t).
> The one place where it is used, you can simply wrap the int32 in an
Immediate().
> 
> 
> On 2010/12/21 01:56:39, Martin Maly wrote:
> > Emit literal zero as xor is used in few places but while browsing through
the
> > code I found that it is not so everywhere.
> > 
> > I ran the benchmarks (crypto, deltablue and friends) with and without my
> change
> > and detected no performance regression.
> > 
> > Thank you!
> > Martin

[William Hesse, 2010-12-22 09:49:54]

LGTM.  I will commit it.