Re: [PATCH] common/mlx5: use intrinsics instead of inline assembly

[Andre Muezerie <andremue@linux.microsoft.com>]

On Tue, Jun 03, 2025 at 05:13:14PM +0200, Dariusz Sosnowski wrote:
> Hi,
> 
> On Mon, May 05, 2025 at 07:57:42AM -0700, Andre Muezerie wrote:
> > When compiling with MSVC the errors below are hit because msvc does not
> > support inline assembly:
> > 
> > 1)
> > ../drivers/common/mlx5/mlx5_common.c(86): warning C4013: '__asm__'
> >     undefined; assuming extern returning int
> > ../drivers/common/mlx5/mlx5_common.c(87): error C2143: syntax error:
> >      missing ')' before ':'
> > 
> > 2)
> > ../drivers/net/mlx5/mlx5_txpp.c(510): error C2065: '__asm__':
> >     undeclared identifier
> > ../drivers/net/mlx5/mlx5_txpp.c(510): error C2143: syntax error:
> >     missing ';' before 'volatile'
> > 
> > The fix for (1) is to use compiler intrinsic __cpuid and for (2)
> > intrinsic _InterlockedCompareExchange128 can be used.
> > 
> > Signed-off-by: Andre Muezerie <andremue@linux.microsoft.com>
> 
> *snip*
> 
> > diff --git a/drivers/net/mlx5/mlx5_txpp.c b/drivers/net/mlx5/mlx5_txpp.c
> > index e6d3ad83e9..5bbaf668e6 100644
> > --- a/drivers/net/mlx5/mlx5_txpp.c
> > +++ b/drivers/net/mlx5/mlx5_txpp.c
> > @@ -486,6 +486,20 @@ mlx5_txpp_cq_arm(struct mlx5_dev_ctx_shared *sh)
> >  }
> >  
> >  #if defined(RTE_ARCH_X86_64)
> > +#ifdef RTE_TOOLCHAIN_MSVC
> > +static inline int
> > +mlx5_atomic128_compare_exchange(rte_int128_t *dst,
> > +				rte_int128_t *exp,
> > +				const rte_int128_t *src)
> > +{
> > +	return (int)_InterlockedCompareExchange128(
> > +		(int64_t volatile *)dst,
> > +		src->val[1], /* exchange high */
> > +		src->val[0], /* exchange low */
> > +		(int64_t *)exp /* comparand result */
> > +	);
> 
> There is one checkpatch warning to fix here:
> 
> 	CHECK:OPEN_ENDED_LINE: Lines should not end with a '('
> 	#117: FILE: drivers/net/mlx5/mlx5_txpp.c:495:
> 	+	return (int)_InterlockedCompareExchange128(
> 
> Also, I don't think that comments for arguments are needed here.
> 
> Could you please organize the code as follows:
> 
> 	return (int)_InterlockedCompareExchange128((int64_t volatile *)dst,
> 		src->val[1], src->val[0], (int64_t *)exp);
> 

Sure. I sent out a v2 of this patch with the suggested change.

> > +}
> > +#else
> >  static inline int
> >  mlx5_atomic128_compare_exchange(rte_int128_t *dst,
> >  				rte_int128_t *exp,
> > @@ -510,6 +524,7 @@ mlx5_atomic128_compare_exchange(rte_int128_t *dst,
> >  	return res;
> >  }
> >  #endif
> > +#endif
> >  
> >  static inline void
> >  mlx5_atomic_read_cqe(rte_int128_t *from, rte_int128_t *ts)
> 
> Best regards,
> Dariusz Sosnowski