Re: [EXT] Re: [PATCH 1/2] config/arm: avoid mcpu and march conflicts

["Juraj Linkeš" <juraj.linkes@pantheon.tech>]

On Mon, Jan 22, 2024 at 12:54 PM Pavan Nikhilesh Bhagavatula
<pbhagavatula@marvell.com> wrote:
>
> > On Sun, Jan 21, 2024 at 10:37 AM <pbhagavatula@marvell.com> wrote:
> > >
> > > From: Pavan Nikhilesh <pbhagavatula@marvell.com>
> > >
> > > The compiler options march and mtune are a subset
> > > of mcpu and will lead to conflicts if improper march
> > > is chosen for a given mcpu.
> > > To avoid conflicts, force part number march when
> > > mcpu is available and is supported by the compiler.
> > >
> > > Example:
> > >         march = armv9-a
> > >         mcpu = neoverse-n2
> > >
> > >         mcpu supported, march supported
> > >         machine_args = ['-mcpu=neoverse-n2', '-march=armv9-a']
> > >
> > >         mcpu supported, march not supported
> > >         machine_args = ['-mcpu=neoverse-n2']
> > >
> > >         mcpu not supported, march supported
> > >         machine_args = ['-march=armv9-a']
> > >
> > >         mcpu not supported, march not supported
> > >         machine_args = ['-march=armv8.6-a']
> > >
> > > Signed-off-by: Pavan Nikhilesh <pbhagavatula@marvell.com>
> > > ---
> > >  config/arm/meson.build | 109 +++++++++++++++++++++++++--------------
> > --
> > >  1 file changed, 67 insertions(+), 42 deletions(-)
> > >
> > > diff --git a/config/arm/meson.build b/config/arm/meson.build
> > > index 36f21d2259..8c8cfccca0 100644
> > > --- a/config/arm/meson.build
> > > +++ b/config/arm/meson.build
> > <snip>
> > > @@ -127,21 +128,22 @@ implementer_cavium = {
> > >      ],
> > >      'part_number_config': {
> > >          '0xa1': {
> > > -            'compiler_options': ['-mcpu=thunderxt88'],
> > > +            'mcpu': 'thunderxt88',
> > >              'flags': flags_part_number_thunderx
> > >          },
> > >          '0xa2': {
> > > -            'compiler_options': ['-mcpu=thunderxt81'],
> > > +            'mcpu': 'thunderxt81',
> > >              'flags': flags_part_number_thunderx
> > >          },
> > >          '0xa3': {
> > > -            'compiler_options': ['-march=armv8-a+crc', '-mcpu=thunderxt83'],
> > > +            'mcpu': 'thunderxt83',
> > > +            'compiler_options': ['-march=armv8-a+crc'],
> >
> > Let's unify this with the rest and specify 'march': 'armv8-a+crc'
> > instead of having it under compiler_options.
>
> Ack.
>
> >
> > >              'flags': flags_part_number_thunderx
> > >          },
> > >          '0xaf': {
> > >              'march': 'armv8.1-a',
> > >              'march_features': ['crc', 'crypto'],
> > > -            'compiler_options': ['-mcpu=thunderx2t99'],
> > > +            'mcpu': 'thunderx2t99',
> > >              'flags': [
> > >                  ['RTE_MACHINE', '"thunderx2"'],
> > >                  ['RTE_ARM_FEATURE_ATOMICS', true],
> > > @@ -153,7 +155,7 @@ implementer_cavium = {
> > >          '0xb2': {
> > >              'march': 'armv8.2-a',
> > >              'march_features': ['crc', 'crypto', 'lse'],
> > > -            'compiler_options': ['-mcpu=octeontx2'],
> > > +            'mcpu': 'octeontx2',
> > >              'flags': [
> > >                  ['RTE_MACHINE', '"cn9k"'],
> > >                  ['RTE_ARM_FEATURE_ATOMICS', true],
> > > @@ -176,7 +178,7 @@ implementer_ampere = {
> > >          '0x0': {
> > >              'march': 'armv8-a',
> > >              'march_features': ['crc', 'crypto'],
> > > -            'compiler_options':  ['-mtune=emag'],
> > > +            'mcpu': 'emag',
> >
> > We're changing mtune to mcpu, is this equivalent?
> >
>
> Both march and mtune are a subset of mcpu.
>

Sure, but we replaced '-mtune=emag' with '-mcpu=emag'. Are these two
builds going to be different or the same?

> > >              'flags': [
> > >                  ['RTE_MACHINE', '"eMAG"'],
> > >                  ['RTE_MAX_LCORE', 32],
> > > @@ -186,7 +188,7 @@ implementer_ampere = {
> > >          '0xac3': {
> > >              'march': 'armv8.6-a',
> > >              'march_features': ['crc', 'crypto'],
> > > -            'compiler_options':  ['-mcpu=ampere1'],
> > > +            'mcpu': 'ampere1',
> > >              'flags': [
> > >                  ['RTE_MACHINE', '"AmpereOne"'],
> > >                  ['RTE_MAX_LCORE', 320],
> > > @@ -206,7 +208,7 @@ implementer_hisilicon = {
> > >          '0xd01': {
> > >              'march': 'armv8.2-a',
> > >              'march_features': ['crypto'],
> > > -            'compiler_options': ['-mtune=tsv110'],
> > > +            'mcpu': 'tsv110',
> > >              'flags': [
> > >                  ['RTE_MACHINE', '"Kunpeng 920"'],
> > >                  ['RTE_ARM_FEATURE_ATOMICS', true],
> > > @@ -695,11 +697,23 @@ if update_flags
> > >
> > >      machine_args = [] # Clear previous machine args
> > >
> > > +    candidate_mcpu = ''
> > > +    support_mcpu = false
> > > +    if part_number_config.has_key('mcpu')
> > > +        mcpu = part_number_config['mcpu']
> > > +        if (cc.has_argument('-mcpu=' + mcpu))
> > > +            candidate_mcpu = mcpu
> > > +            support_mcpu = true
> > > +        endif
> > > +    endif
> > > +
> > >      # probe supported archs and their features
> > >      candidate_march = ''
> > >      if part_number_config.has_key('march')
> > > -        if part_number_config.get('force_march', false)
> > > -            candidate_march = part_number_config['march']
> > > +        if part_number_config.get('force_march', false) or support_mcpu
> >
> > Instead of using the extra "support_mcpu" variable, we could do the
> > same check as with candidate march (if candidate_mcpu != '', which we
> > actually do below in the last lines of the patch).
> >
>
> Ack.
>
> > If I understand the logic correctly, we don't want to do the march
> > fallback if mcpu is specified - either the march works with the given
> > mcpu or we do without it (because we don't actually need it with
> > mcpu). Is that correct?
> >
>
> Yes, but still exact march defined in part_number_config should be present for setting extra_march_features.
> specially for expressing crypto support.
>

Ok, thanks.

> > > +            if cc.has_argument('-march=' +  part_number_config['march'])
> >
> > Now that we've added mcpu into the mix, is this still the right
> > condition? Can the below happen?
> >
> > This check finds that machine_args = ['-march=armv9-a'] is supported.
> >
> > But taken together with mcpu (machine_args = ['-mcpu=neoverse-n2',
> > '-march=armv9-a']), it is not supported? In this case we'll end up
> > with invalid configuration.
>
> This is the only correct option and evolves into -march=armv9-a+sve2+crypto for cn10k
> whereas other neoverse-n2 might only have -march=armv9-a+sve2.
>

Maybe I should rephrase my question a bit:

The correct options are ['-mcpu=neoverse-n2', '-march=armv9-a']. Is it
possible that the compiler will say:

['-mcpu=neoverse-n2', '-march=armv9-a'] is supported
['-mcpu=neoverse-n2'] is supported
['-march=armv9-a'] is not supported

So basically the question is are we risking that the compiler will say
it supports both options only when both are passed while also saying
it doesn't support one or both of them when checked alone. We've seen
this behavior with newer compilers in aarch32 builds
(-march=armv8-a+simd -mfpu=auto are supported when both are passed,
but -march=armv8-a is not supported alone), so I wanted to be sure.

> Example:
>
> Good:
> #aarch64-linux-gnu-gcc  -march=armv9-a+sve2+crypto -mcpu=neoverse-n2 shrn.c
> #aarch64-linux-gnu-gcc  -march=armv9-a+sve2 -mcpu=neoverse-n2 shrn.c
> #aarch64-linux-gnu-gcc  -march=armv9-a -mcpu=neoverse-n2 shrn.c
>
> Bad:
> #aarch64-linux-gnu-gcc -mcpu=neoverse-n2 -march=armv8-a shrn.c
> cc1: warning: switch '-mcpu=neoverse-n2' conflicts with '-march=armv8-a' switch
> #aarch64-linux-gnu-gcc -mcpu=neoverse-n2 -march=armv8.1-a shrn.c
> cc1: warning: switch '-mcpu=neoverse-n2' conflicts with '-march=armv8.1-a' switch
> #aarch64-linux-gnu-gcc -mcpu=neoverse-n2 -march=armv8.2-a shrn.c
> cc1: warning: switch '-mcpu=neoverse-n2' conflicts with '-march=armv8.2-a' switch
> #aarch64-linux-gnu-gcc -mcpu=neoverse-n2 -march=armv8.3-a shrn.c
> cc1: warning: switch '-mcpu=neoverse-n2' conflicts with '-march=armv8.3-a' switch
> #aarch64-linux-gnu-gcc -mcpu=neoverse-n2 -march=armv8.4-a shrn.c
> cc1: warning: switch '-mcpu=neoverse-n2' conflicts with '-march=armv8.4-a' switch
> #aarch64-linux-gnu-gcc -mcpu=neoverse-n2 -march=armv8.5-a shrn.c
> cc1: warning: switch '-mcpu=neoverse-n2' conflicts with '-march=armv8.5-a' switch
> #aarch64-linux-gnu-gcc -mcpu=neoverse-n2 -march=armv8.6-a shrn.c
> cc1: warning: switch '-mcpu=neoverse-n2' conflicts with '-march=armv8.6-a' switch
> #aarch64-linux-gnu-gcc -mcpu=neoverse-n2 -march=armv8.7-a shrn.c
> cc1: warning: switch '-mcpu=neoverse-n2' conflicts with '-march=armv8.7-a' switch
>
> >
> > > +                candidate_march = part_number_config['march']
> > > +            endif
> > >          else
> > >              supported_marchs = ['armv8.6-a', 'armv8.5-a', 'armv8.4-a', 'armv8.3-
> > a',
> > >                                  'armv8.2-a', 'armv8.1-a', 'armv8-a']