Re: [PATCH v2 1/4] perf build: enable -fno-strict-aliasing

[Eric Biggers]

On Sun, Jun 15, 2025 at 04:40:45PM -0700, Ian Rogers wrote:
> On Fri, Jun 13, 2025 at 9:43 PM Eric Biggers <ebiggers@xxxxxxxxxx> wrote:
> >
> > From: Eric Biggers <ebiggers@xxxxxxxxxx>
> >
> > perf pulls in code from kernel headers that assumes it is being built
> > with -fno-strict-aliasing, namely put_unaligned_*() from
> > <linux/unaligned.h> which write the data using packed structs that lack
> > the may_alias attribute.  Enable -fno-strict-aliasing to prevent
> > miscompilations in sha1.c which would otherwise occur due to this issue.
> 
> Wow, good catch! I wonder if -fsanitize=type could be used to capture
> when perf's code is broken like this? Perhaps we should just remove
> linux/unaligned.h in tools because of this, the alternative of using
> memcpy doesn't look particularly burdensome. Given the memcpys are of
> a known/fixed size I'd expect the compiler to be able to optimize
> things just as well. Perhaps we should rewrite unaligned.h in tools
> but perhaps the kernel too. Something like:
> 
> #define __get_unaligned_t(type, ptr) ({
>                  \
>         const struct { type x; } __packed * __get_pptr =
> (typeof(__get_pptr))(ptr);     \
>         __get_pptr->x;
>                  \
> })
> 
> becomes:
> 
> #define __get_unaligned_t(type, ptr) ({
>                  \
>         type __get_val; memcpy(&__get_val, ptr, sizeof(__get_val));     \
>         __get_val;
>              \
> })

As far as I know, the packed struct method of doing unaligned memory accesses is
obsolete these days, and memcpy() generates the desired code on all supported
architectures and compilers.

- Eric