Re: [dpdk-dev] [PATCH] ixgbe: fix clang compile - remove truncation errors

[Bruce Richardson <bruce.richardson@intel.com>]

On Mon, Dec 01, 2014 at 10:18:06AM -0500, Neil Horman wrote:
> On Mon, Dec 01, 2014 at 02:36:46PM +0000, Bruce Richardson wrote:
> > On Mon, Dec 01, 2014 at 09:25:44AM -0500, Neil Horman wrote:
> > > On Mon, Dec 01, 2014 at 11:24:58AM +0000, Bruce Richardson wrote:
> > > > On Mon, Dec 01, 2014 at 06:18:17AM -0500, Neil Horman wrote:
> > > > > On Mon, Dec 01, 2014 at 10:09:38AM +0100, Olivier MATZ wrote:
> > > > > > Hi Bruce, Hi Neil,
> > > > > > 
> > > > > > On 11/30/2014 02:05 AM, Neil Horman wrote:
> > > > > > > On Fri, Nov 28, 2014 at 03:31:00PM +0000, Bruce Richardson wrote:
> > > > > > >> When compiling with clang, errors were being emitted due to truncation
> > > > > > >> of values when assigning to the tx_offload_mask bit fields.
> > > > > > >>
> > > > > > >> dpdk.org/lib/librte_pmd_ixgbe/ixgbe_rxtx.c:404:27: fatal error: implicit truncation from 'int' to bitfield changes value from -1 to 127 [-Wbitfield-constant-conversion]
> > > > > > >> 		    tx_offload_mask.l2_len = ~0;
> > > > > > >>
> > > > > > >> The fix proposed here is to define a static const value of the same type
> > > > > > >> with all fields set to 1s, and use that instead of constants for assigning to.
> > > > > > >>
> > > > > > >> Other options would be to explicitily define the suitable constants that
> > > > > > >> would not truncate for each individual field e.g. 0x7f for l2_len, 0x1FF
> > > > > > >> for l3_len, etc., but this solution here has the advantage that it works
> > > > > > >> without any changes to values if the field sizes are ever modified.
> > > > > > >>
> > > > > > >> Signed-off-by: Bruce Richardson <bruce.richardson@intel.com>
> > > > > > >> ---
> > > > > > >>  lib/librte_pmd_ixgbe/ixgbe_rxtx.c | 29 +++++++++++++++--------------
> > > > > > >>  1 file changed, 15 insertions(+), 14 deletions(-)
> > > > > > >>
> > > > > > >> diff --git a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > > >> index 8559ef6..4f71194 100644
> > > > > > >> --- a/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > > >> +++ b/lib/librte_pmd_ixgbe/ixgbe_rxtx.c
> > > > > > >> @@ -367,6 +367,7 @@ ixgbe_set_xmit_ctx(struct igb_tx_queue* txq,
> > > > > > >>  		volatile struct ixgbe_adv_tx_context_desc *ctx_txd,
> > > > > > >>  		uint64_t ol_flags, union ixgbe_tx_offload tx_offload)
> > > > > > >>  {
> > > > > > >> +	static const union ixgbe_tx_offload offload_allones = { .data = ~0 };
> > > > > > > Do you want to make this a static data structure?  If you make it a macro like
> > > > > > > this:
> > > > > > > #define ALLONES {.data = ~0}
> > > > > > > Then you save the extra data space in the .data area (not that its that much),
> > > > > > > and you can define it in a header file and use it in multiple c files (if you
> > > > > > > need to)
> > > > > > 
> > > > > > I found that the following code works:
> > > > > > 
> > > > > > 	tx_offload_mask.l2_len |= ~0;
> > > > > > 
> > > > > > (note the '|=' instead of '=')
> > > > > > 
> > > > > How exactly does this work? does the or operator imply some level of type
> > > > > promotion that the assignment doesn't to avoid the truncation?
> > > > > Neil
> > > > > 
> > > > 
> > > > For any aithmetic, and presumably logical, operation on two values, any values
> > > > smaller than "int" are promoted to type int before the operation takes place. I
> > > > believe the exact rules for this are in the C specs e.g. C99.
> > > > 
> > > Yes, but I would have thought that assignment was included in the list of
> > > logical operations for that promotion to occur.  The above change seems to
> > > suggest it isn't, which is why I'm asking.  C99 specifies |= explicitly as a
> > > type of assignment operator (see 6.5.16 here:
> > > http://www.open-std.org/jtc1/sc22/WG14/www/docs/n1256.pdf
> > > )
> > > 
> > > So I would presume that using = should work exactly the same as |= in terms of
> > > type promotion.  We also don't get this warning on gcc, which concerns me that
> > > we might just be papering over a compiler problem here.
> > > 
> > > Looking at the error, its complaining that we're truncating an int value to a
> > > bitfield (which we are), and that the resultant value is 127 rather that -1
> > > (which it would be if we actually intended to treat it as an integer
> > > 
> > > Which I think is where the problem lies.  That is to say we've typed the
> > > bitfields in ixgbe_tx_offload as uint64_t.  I'm guessing gcc just promotes ~0 to
> > > an unsigned int during the assignment, and supresses the warning (unless you
> > > turn on -pedantic or some such), but clang does not.  The correct solution I
> > > think here is to either:
> > > 
> > > 1) modify the bitfield types so that they are signed integers, thereby
> > > maintaining the resultant value of -1 after the assignment
> > > 
> > > or
> > > 
> > > 2) Modify the ~0 to be ~0UL, so that the clang compiler sees that the resultant
> > > value will be MAXLONG after the assignment
> > > 
> > > I'd think operation 2 would be the better choice
> > > Neil
> > >
> > I'm not a compiler expert, but looking at it a bit more what I think is 
> > happening is that we are simply changing the assignment from a constant one to
> > a computed one instead. With the constant assignment, the compiler can check that
> > the assignment doesn't overflow, while with the computed value, it has no choice
> > to accept the truncation since any computation is going to take place with variables
> > of at least size "int" and there is no way to typecast the resulting value to
> > a bit field.
> > 
> > As for papering-over compiler niggles, possibly so, but this solution is shorter and
> > less impactful than the other solutions which are less workaround-like - i.e. those
> > that assign values of exactly the right size using either magic numbers or a
> > special-value copy of the structure.
> > 
> > Also, in terms of the two options you propose, I tried the second and it still gives
> > errors, so the signed-ness or unsigned-ness is not the problem the compiler has, its
> > the truncation.
> > 
> >   CC ixgbe_rxtx.o
> >   /usr/home/bruce/dpdk.org/lib/librte_pmd_ixgbe/ixgbe_rxtx.c:384:28: fatal error: implicit truncation from 'unsigned long' to bitfield changes value from 18446744073709551615 to 65535
> >         [-Wbitfield-constant-conversion]
> > 	                tx_offload_mask.vlan_tci = ~0UL;
> > 
> > 
> > /Bruce
> > 
> You're right, it does, and looking at it now, I wonder if the warning really
> just needs to be removed.  Gcc has no such warning in it currently (though I
> expect -pedantic would say something here).  Regardless, looking at the clang
> docs I can't find any documentation about the bitfield-constant-conversion
> warning, and it seems to exist only to tell us that we're truncating an integer
> to an integer of a smaller size (which will clearly be the case anytime we are
> assigning a constant to a bitfield).  Instead of avoiding the warning by doing
> any sort of code trickery, why not just remove the warning?
> 
> Neil
>
That's one for a community discussion. My opinion on it is that disabling warnings
is the last option I'd look for - as a point of principle. I'd prefer to take 
Olivier's fix, or my proposed fix, over disabling the warnings. The fix is fairly
trivial, and I think disabling warnings for something like this would set a bad
precedent.

/Bruce