Hi,

On Fri, Jun 24, 2016 at 01:41:05PM -0700, Evgeny Stupachenko wrote:
> Hi,
> 
> Fix ICE when IPA-cp and target_clones are applied to the same function.
> Is the patch ok for trunk?

I can't approve anything but since I wrote most of IPA-CP, it may
count that I am fine with the patch.

However, it should also be backported to the 6 branch.

In the future, it is useful to file a bugzilla PR for issues like
this, even if you also provide a fix.  It helps with tracking
backports and with a PR, you would have probably caught my attention
sooner.

In any event, thanks for addressing this.

Martin

> 
> Thanks,
> Evgeny
> 
> 2016-06-24  Evgeny Stupachenko  <evstu...@gmail.com>
> 
> gcc/
>         * ipa-cp.c (determine_versionability): Do not create constprop clones,
>         when target_clones attribute is set.
> diff --git a/gcc/ipa-cp.c b/gcc/ipa-cp.c
> index 2710494..4b642ba 100644
> --- a/gcc/ipa-cp.c
> +++ b/gcc/ipa-cp.c
> @@ -533,6 +533,13 @@ determine_versionability (struct cgraph_node *node,
>          coexist, but that may not be worth the effort.  */
>        reason = "function has SIMD clones";
>      }
> +  else if (lookup_attribute ("target_clones", DECL_ATTRIBUTES (node->decl)))
> +    {
> +      /* Ideally we should clone the target clones themselves and create
> +        copies of them, so IPA-cp and target clones can happily
> +        coexist, but that may not be worth the effort.  */
> +      reason = "function target_clones attribute";
> +    }
>    /* Don't clone decls local to a comdat group; it breaks and for C++
>      decloned constructors, inlining is always better anyway.  */
>    else if (node->comdat_local_p ())
> diff --git a/gcc/testsuite/gcc.target/i386/mvc8.c
> b/gcc/testsuite/gcc.target/i386/mvc8.c
> new file mode 100644
> index 0000000..e9ab9e1
> --- /dev/null
> +++ b/gcc/testsuite/gcc.target/i386/mvc8.c
> @@ -0,0 +1,14 @@
> +/* { dg-do compile } */
> +/* { dg-require-ifunc "" } */
> +/* { dg-options "-O3 -fno-inline" } */
> +/* { dg-final { scan-assembler-not "constprop" } } */
> +__attribute__((target_clones("arch=core-avx2","arch=slm","default")))
> +void foo (float *a, int b) {
> +    *a = (float)b;
> +}
> +float a;
> +int main() {
> +  int i;
> +  for (i = 0; i < 1024; i++)
> +    foo (&a, 5);
> +}

Reply via email to