[dpdk-dev] [PATCH] test-pmd: Fix pointer aliasing error

[Bruce Richardson]

On Wed, Dec 03, 2014 at 03:19:34PM +0000, Qiu, Michael wrote:
> On 2014/12/3 22:51, Richardson, Bruce wrote:
> > On Wed, Dec 03, 2014 at 01:59:58PM +0000, Qiu, Michael wrote:
> >> On 2014/12/3 19:43, Richardson, Bruce wrote:
> >>> On Wed, Dec 03, 2014 at 07:28:19PM +0800, Michael Qiu wrote:
> >>>> app/test-pmd/csumonly.c: In function ‘get_psd_sum’:
> >>>> build/include/rte_ip.h:161: error: dereferencing pointer ‘u16’
> >>>> 	does break strict-aliasing rules
> >>>> build/include/rte_ip.h:157: note: initialized from here
> >>>> 	...
> >>>>
> >>>> The root cause is that, compile enable strict aliasing by default,
> >>>> while in function rte_raw_cksum() try to convert 'const char *'
> >>>> to 'const uint16_t *'.
> >>>>
> >>> What compiler version is this with? Is there any other way to fix this
> >>> other than disabling the compiler warnings. Turning off strict aliasing may
> >>> affect performance as it reduces the number of optimizations that the compiler
> >>> can perform.
> >> The compile version is:
> >>
> >> $ gcc -v
> >> Using built-in specs.
> >> Target: x86_64-redhat-linux
> >> ...
> >> gcc version 4.4.7 20120313 (Red Hat 4.4.7-4) (GCC)
> >>
> >>
> >> The OS is centos6.5 x86_64
> >>
> >>
> >> Actually, another possible solution is, as gcc manual shows, use union.
> >> In function rte_raw_cksum() of lib/librte_net/rte_ip.h:
> >>
> >> static inline uint16_t
> >> rte_raw_cksum(const char *buf, size_t len){
> >>     union {
> >>         const char *ubuf;
> >>         const uint16_t *uu16;
> >>     } convert;
> >>
> >>     convert.ubuf = buf;
> >>     const uint16_t *u16 = convert.uu16;
> >>     ...
> >> }
> >>
> >> This may be work, but not test yet.
> >>
> >> Any comments about this solution?
> > what happens if you make rte_raw_cksum take a void * (or const void *) parameter
> > instead of "const char *"?
> 
> "size_t len" is for type char, it is the the array size(for char array
> is byte numbers), if we use void *, the meaning maybe confuse I think.
That shouldn't be a big issue. We can rename it to "size" instead of "len" if
needed.

> But it should work with other code change.
> 
> Thanks,
> Michael
> >
> > /Bruce
> >
>