Penitum II & egcs 1.1.1 oddity

Penitum II & egcs 1.1.1 oddity

[Barry Shilliday]

I've compiled egcs 1.1.1 having tried the standard 1.1.1 release and the
1.1.1 patched release.  I've compiled on RedHat Linux 5.2 / PII 400.  I
compiled using 'configure --enable-shared --enable-haifa'.

Is it normal that the compiler is sending -mcpu=pentiumpro but
-march=pentium by default?  For example, this is the output from 'gcc
-fverbose-asm -S test.c'

# GNU C version egcs-2.91.60.1 19990115/Linux (egcs-1.1.1 release)
(i686-pc-linux-gnu) compiled by GNU C version egcs-2.91.60.1
19990115/Linux (egcs-1.1.1 release).
# options passed:  -fverbose-asm
# options enabled:  -fpeephole -ffunction-cse -fkeep-static-consts
# -fpcc-struct-return -fsched-interblock -fsched-spec -fcommon
# -fverbose-asm -fgnu-linker -fargument-alias -m80387 -mhard-float
# -mno-soft-float -mieee-fp -mfp-ret-in-387 -mschedule-prologue
# -mcpu=pentiumpro -march=pentium

Does it make a lot of difference compiling using -march=pentium as
opposed to -march=pentiumpro on a Pentium II?

Any comments appreciated

-- 
Barry Shilliday

Re: Penitum II & egcs 1.1.1 oddity

[Jeffrey A Law]

  In message < 36A85F12.653EF986@nibelung.demon.co.uk >you write:
  > I've compiled egcs 1.1.1 having tried the standard 1.1.1 release and the
  > 1.1.1 patched release.  I've compiled on RedHat Linux 5.2 / PII 400.  I
  > compiled using 'configure --enable-shared --enable-haifa'.
  > 
  > Is it normal that the compiler is sending -mcpu=pentiumpro but
  > -march=pentium by default?  For example, this is the output from 'gcc
  > -fverbose-asm -S test.c'
Yes, this is normal.

It means schedule for a ppro, but generate (effectively) common-mode
instructions (ie your code can run on all x86 variants).


  > Does it make a lot of difference compiling using -march=pentium as
  > opposed to -march=pentiumpro on a Pentium II?
"a lot" is a relative term.

-march=pentiumpro will enable ppro specific instructions, namely conditional
moves, which for some code can be a nice improvement.  For other code it won't
help in any noticable way.  Just depends on your application.

jeff

RE: Penitum II & egcs 1.1.1 oddity

[Barry]

>   > Does it make a lot of difference compiling using -march=pentium as
>   > opposed to -march=pentiumpro on a Pentium II?
> "a lot" is a relative term.
>
> -march=pentiumpro will enable ppro specific instructions, namely
> conditional
> moves, which for some code can be a nice improvement.  For other
> code it won't
> help in any noticable way.  Just depends on your application.
>
Thanks for the information.  I'd be correct in assuming
that -march=pentiumpro would never cause a loss of performance
over -march=pentium?

--
Barry

Re: Penitum II & egcs 1.1.1 oddity

[Jeffrey A Law]

  In message < 000201be4624$e16f7e60$71aa989e@nibelung >you write:
  > >   > Does it make a lot of difference compiling using -march=pentium as
  > >   > opposed to -march=pentiumpro on a Pentium II?
  > > "a lot" is a relative term.
  > >
  > > -march=pentiumpro will enable ppro specific instructions, namely
  > > conditional
  > > moves, which for some code can be a nice improvement.  For other
  > > code it won't
  > > help in any noticable way.  Just depends on your application.
  > >
  > Thanks for the information.  I'd be correct in assuming
  > that -march=pentiumpro would never cause a loss of performance
  > over -march=pentium?
"never" is an extremely strong word.  Given time I could probably come up
with cases where cmoves could hurt.  For example, by lengthening a basic block
we could expose more loop invariants to the loop optimizer, which would pull
them out the loop, increasing register pressure inside the loop, which could
calls spills in the loop and possibly a net loss in performance.

Having said that, I would expect any loss of performance due to using that
option to be rare relative to the performance gains.


jeff

Re: Penitum II & egcs 1.1.1 oddity

[Jeffrey A Law]

  In message < 36A85F12.653EF986@nibelung.demon.co.uk >you write:
  > I've compiled egcs 1.1.1 having tried the standard 1.1.1 release and the
  > 1.1.1 patched release.  I've compiled on RedHat Linux 5.2 / PII 400.  I
  > compiled using 'configure --enable-shared --enable-haifa'.
  > 
  > Is it normal that the compiler is sending -mcpu=pentiumpro but
  > -march=pentium by default?  For example, this is the output from 'gcc
  > -fverbose-asm -S test.c'
Yes, this is normal.

It means schedule for a ppro, but generate (effectively) common-mode
instructions (ie your code can run on all x86 variants).


  > Does it make a lot of difference compiling using -march=pentium as
  > opposed to -march=pentiumpro on a Pentium II?
"a lot" is a relative term.

-march=pentiumpro will enable ppro specific instructions, namely conditional
moves, which for some code can be a nice improvement.  For other code it won't
help in any noticable way.  Just depends on your application.

jeff

Re: Penitum II & egcs 1.1.1 oddity

[Jeffrey A Law]

  In message < 000201be4624$e16f7e60$71aa989e@nibelung >you write:
  > >   > Does it make a lot of difference compiling using -march=pentium as
  > >   > opposed to -march=pentiumpro on a Pentium II?
  > > "a lot" is a relative term.
  > >
  > > -march=pentiumpro will enable ppro specific instructions, namely
  > > conditional
  > > moves, which for some code can be a nice improvement.  For other
  > > code it won't
  > > help in any noticable way.  Just depends on your application.
  > >
  > Thanks for the information.  I'd be correct in assuming
  > that -march=pentiumpro would never cause a loss of performance
  > over -march=pentium?
"never" is an extremely strong word.  Given time I could probably come up
with cases where cmoves could hurt.  For example, by lengthening a basic block
we could expose more loop invariants to the loop optimizer, which would pull
them out the loop, increasing register pressure inside the loop, which could
calls spills in the loop and possibly a net loss in performance.

Having said that, I would expect any loss of performance due to using that
option to be rare relative to the performance gains.


jeff

RE: Penitum II & egcs 1.1.1 oddity

[Barry]

>   > Does it make a lot of difference compiling using -march=pentium as
>   > opposed to -march=pentiumpro on a Pentium II?
> "a lot" is a relative term.
>
> -march=pentiumpro will enable ppro specific instructions, namely
> conditional
> moves, which for some code can be a nice improvement.  For other
> code it won't
> help in any noticable way.  Just depends on your application.
>
Thanks for the information.  I'd be correct in assuming
that -march=pentiumpro would never cause a loss of performance
over -march=pentium?

--
Barry

Penitum II & egcs 1.1.1 oddity

[Barry Shilliday]

I've compiled egcs 1.1.1 having tried the standard 1.1.1 release and the
1.1.1 patched release.  I've compiled on RedHat Linux 5.2 / PII 400.  I
compiled using 'configure --enable-shared --enable-haifa'.

Is it normal that the compiler is sending -mcpu=pentiumpro but
-march=pentium by default?  For example, this is the output from 'gcc
-fverbose-asm -S test.c'

# GNU C version egcs-2.91.60.1 19990115/Linux (egcs-1.1.1 release)
(i686-pc-linux-gnu) compiled by GNU C version egcs-2.91.60.1
19990115/Linux (egcs-1.1.1 release).
# options passed:  -fverbose-asm
# options enabled:  -fpeephole -ffunction-cse -fkeep-static-consts
# -fpcc-struct-return -fsched-interblock -fsched-spec -fcommon
# -fverbose-asm -fgnu-linker -fargument-alias -m80387 -mhard-float
# -mno-soft-float -mieee-fp -mfp-ret-in-387 -mschedule-prologue
# -mcpu=pentiumpro -march=pentium

Does it make a lot of difference compiling using -march=pentium as
opposed to -march=pentiumpro on a Pentium II?

Any comments appreciated

-- 
Barry Shilliday

RE: Penitum II & egcs 1.1.1 oddity

[Barry]

> << Does it make a lot of difference compiling using -march=pentium as
>  opposed to -march=pentiumpro on a Pentium II? >>
> Unlikely to make much difference, although there appears to be a latent
> problem with -march=pentiumpro in i386.md which Jeff Law just
> diagnosed this
> week.  It's never bitten me with egcs-1.1.1 but you never know.

Thanks for the information.  I'll modify the spec file to
set -march=pentiumpro by default for the time being.

--
Barry

Re: Penitum II & egcs 1.1.1 oddity

[N8TM]

In a message dated 1/22/99 3:23:54 AM Pacific Standard Time,
barry@nibelung.demon.co.uk writes:

<< Does it make a lot of difference compiling using -march=pentium as
 opposed to -march=pentiumpro on a Pentium II? >>
Unlikely to make much difference, although there appears to be a latent
problem with -march=pentiumpro in i386.md which Jeff Law just diagnosed this
week.  It's never bitten me with egcs-1.1.1 but you never know.

RE: Penitum II & egcs 1.1.1 oddity

[Barry]

> << Does it make a lot of difference compiling using -march=pentium as
>  opposed to -march=pentiumpro on a Pentium II? >>
> Unlikely to make much difference, although there appears to be a latent
> problem with -march=pentiumpro in i386.md which Jeff Law just
> diagnosed this
> week.  It's never bitten me with egcs-1.1.1 but you never know.

Thanks for the information.  I'll modify the spec file to
set -march=pentiumpro by default for the time being.

--
Barry

Re: Penitum II & egcs 1.1.1 oddity

[N8TM]

In a message dated 1/22/99 3:23:54 AM Pacific Standard Time,
barry@nibelung.demon.co.uk writes:

<< Does it make a lot of difference compiling using -march=pentium as
 opposed to -march=pentiumpro on a Pentium II? >>
Unlikely to make much difference, although there appears to be a latent
problem with -march=pentiumpro in i386.md which Jeff Law just diagnosed this
week.  It's never bitten me with egcs-1.1.1 but you never know.